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);
447 /// isAddress - Returns true if the specified instruction is using the
448 /// specified value as an address.
449 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
450 bool isAddress = isa<LoadInst>(Inst);
451 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
452 if (SI->getOperand(1) == OperandVal)
454 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
455 // Addressing modes can also be folded into prefetches and a variety
457 switch (II->getIntrinsicID()) {
459 case Intrinsic::prefetch:
460 case Intrinsic::x86_sse2_loadu_dq:
461 case Intrinsic::x86_sse2_loadu_pd:
462 case Intrinsic::x86_sse_loadu_ps:
463 case Intrinsic::x86_sse_storeu_ps:
464 case Intrinsic::x86_sse2_storeu_pd:
465 case Intrinsic::x86_sse2_storeu_dq:
466 case Intrinsic::x86_sse2_storel_dq:
467 if (II->getOperand(1) == OperandVal)
475 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
476 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
477 /// return true. Otherwise, return false.
478 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
479 SmallPtrSet<Instruction*,16> &Processed) {
480 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
481 return false; // Void and FP expressions cannot be reduced.
482 if (!Processed.insert(I))
483 return true; // Instruction already handled.
485 // Get the symbolic expression for this instruction.
486 SCEVHandle ISE = GetExpressionSCEV(I);
487 if (isa<SCEVCouldNotCompute>(ISE)) return false;
489 // Get the start and stride for this expression.
490 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
491 SCEVHandle Stride = Start;
492 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
493 return false; // Non-reducible symbolic expression, bail out.
495 std::vector<Instruction *> IUsers;
496 // Collect all I uses now because IVUseShouldUsePostIncValue may
497 // invalidate use_iterator.
498 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
499 IUsers.push_back(cast<Instruction>(*UI));
501 for (unsigned iused_index = 0, iused_size = IUsers.size();
502 iused_index != iused_size; ++iused_index) {
504 Instruction *User = IUsers[iused_index];
506 // Do not infinitely recurse on PHI nodes.
507 if (isa<PHINode>(User) && Processed.count(User))
510 // If this is an instruction defined in a nested loop, or outside this loop,
511 // don't recurse into it.
512 bool AddUserToIVUsers = false;
513 if (LI->getLoopFor(User->getParent()) != L) {
514 DOUT << "FOUND USER in other loop: " << *User
515 << " OF SCEV: " << *ISE << "\n";
516 AddUserToIVUsers = true;
517 } else if (!AddUsersIfInteresting(User, L, Processed)) {
518 DOUT << "FOUND USER: " << *User
519 << " OF SCEV: " << *ISE << "\n";
520 AddUserToIVUsers = true;
523 if (AddUserToIVUsers) {
524 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
525 if (StrideUses.Users.empty()) // First occurance of this stride?
526 StrideOrder.push_back(Stride);
528 // Okay, we found a user that we cannot reduce. Analyze the instruction
529 // and decide what to do with it. If we are a use inside of the loop, use
530 // the value before incrementation, otherwise use it after incrementation.
531 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
532 // The value used will be incremented by the stride more than we are
533 // expecting, so subtract this off.
534 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
535 StrideUses.addUser(NewStart, User, I);
536 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
537 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
539 StrideUses.addUser(Start, User, I);
547 /// BasedUser - For a particular base value, keep information about how we've
548 /// partitioned the expression so far.
550 /// SE - The current ScalarEvolution object.
553 /// Base - The Base value for the PHI node that needs to be inserted for
554 /// this use. As the use is processed, information gets moved from this
555 /// field to the Imm field (below). BasedUser values are sorted by this
559 /// Inst - The instruction using the induction variable.
562 /// OperandValToReplace - The operand value of Inst to replace with the
564 Value *OperandValToReplace;
566 /// Imm - The immediate value that should be added to the base immediately
567 /// before Inst, because it will be folded into the imm field of the
571 // isUseOfPostIncrementedValue - True if this should use the
572 // post-incremented version of this IV, not the preincremented version.
573 // This can only be set in special cases, such as the terminating setcc
574 // instruction for a loop and uses outside the loop that are dominated by
576 bool isUseOfPostIncrementedValue;
578 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
579 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
580 OperandValToReplace(IVSU.OperandValToReplace),
581 Imm(SE->getIntegerSCEV(0, Base->getType())),
582 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
584 // Once we rewrite the code to insert the new IVs we want, update the
585 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
587 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
588 Instruction *InsertPt,
589 SCEVExpander &Rewriter, Loop *L, Pass *P,
590 SmallVectorImpl<Instruction*> &DeadInsts);
592 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
593 SCEVExpander &Rewriter,
594 Instruction *IP, Loop *L);
599 void BasedUser::dump() const {
600 cerr << " Base=" << *Base;
601 cerr << " Imm=" << *Imm;
602 cerr << " Inst: " << *Inst;
605 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
606 SCEVExpander &Rewriter,
607 Instruction *IP, Loop *L) {
608 // Figure out where we *really* want to insert this code. In particular, if
609 // the user is inside of a loop that is nested inside of L, we really don't
610 // want to insert this expression before the user, we'd rather pull it out as
611 // many loops as possible.
612 LoopInfo &LI = Rewriter.getLoopInfo();
613 Instruction *BaseInsertPt = IP;
615 // Figure out the most-nested loop that IP is in.
616 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
618 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
619 // the preheader of the outer-most loop where NewBase is not loop invariant.
620 if (L->contains(IP->getParent()))
621 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
622 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
623 InsertLoop = InsertLoop->getParentLoop();
626 // If there is no immediate value, skip the next part.
628 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
630 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
632 // If we are inserting the base and imm values in the same block, make sure to
633 // adjust the IP position if insertion reused a result.
634 if (IP == BaseInsertPt)
635 IP = Rewriter.getInsertionPoint();
637 // Always emit the immediate (if non-zero) into the same block as the user.
638 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
639 return Rewriter.expandCodeFor(NewValSCEV, IP);
644 // Once we rewrite the code to insert the new IVs we want, update the
645 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
646 // to it. NewBasePt is the last instruction which contributes to the
647 // value of NewBase in the case that it's a diffferent instruction from
648 // the PHI that NewBase is computed from, or null otherwise.
650 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
651 Instruction *NewBasePt,
652 SCEVExpander &Rewriter, Loop *L, Pass *P,
653 SmallVectorImpl<Instruction*> &DeadInsts){
654 if (!isa<PHINode>(Inst)) {
655 // By default, insert code at the user instruction.
656 BasicBlock::iterator InsertPt = Inst;
658 // However, if the Operand is itself an instruction, the (potentially
659 // complex) inserted code may be shared by many users. Because of this, we
660 // want to emit code for the computation of the operand right before its old
661 // computation. This is usually safe, because we obviously used to use the
662 // computation when it was computed in its current block. However, in some
663 // cases (e.g. use of a post-incremented induction variable) the NewBase
664 // value will be pinned to live somewhere after the original computation.
665 // In this case, we have to back off.
667 // If this is a use outside the loop (which means after, since it is based
668 // on a loop indvar) we use the post-incremented value, so that we don't
669 // artificially make the preinc value live out the bottom of the loop.
670 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
671 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
672 InsertPt = NewBasePt;
674 } else if (Instruction *OpInst
675 = dyn_cast<Instruction>(OperandValToReplace)) {
677 while (isa<PHINode>(InsertPt)) ++InsertPt;
680 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
681 // Adjust the type back to match the Inst. Note that we can't use InsertPt
682 // here because the SCEVExpander may have inserted the instructions after
683 // that point, in its efforts to avoid inserting redundant expressions.
684 if (isa<PointerType>(OperandValToReplace->getType())) {
685 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
687 OperandValToReplace->getType());
689 // Replace the use of the operand Value with the new Phi we just created.
690 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
691 DOUT << " CHANGED: IMM =" << *Imm;
692 DOUT << " \tNEWBASE =" << *NewBase;
693 DOUT << " \tInst = " << *Inst;
697 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
698 // expression into each operand block that uses it. Note that PHI nodes can
699 // have multiple entries for the same predecessor. We use a map to make sure
700 // that a PHI node only has a single Value* for each predecessor (which also
701 // prevents us from inserting duplicate code in some blocks).
702 DenseMap<BasicBlock*, Value*> InsertedCode;
703 PHINode *PN = cast<PHINode>(Inst);
704 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
705 if (PN->getIncomingValue(i) == OperandValToReplace) {
706 // If this is a critical edge, split the edge so that we do not insert the
707 // code on all predecessor/successor paths. We do this unless this is the
708 // canonical backedge for this loop, as this can make some inserted code
709 // be in an illegal position.
710 BasicBlock *PHIPred = PN->getIncomingBlock(i);
711 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
712 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
714 // First step, split the critical edge.
715 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
717 // Next step: move the basic block. In particular, if the PHI node
718 // is outside of the loop, and PredTI is in the loop, we want to
719 // move the block to be immediately before the PHI block, not
720 // immediately after PredTI.
721 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
722 BasicBlock *NewBB = PN->getIncomingBlock(i);
723 NewBB->moveBefore(PN->getParent());
726 // Splitting the edge can reduce the number of PHI entries we have.
727 e = PN->getNumIncomingValues();
730 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
732 // Insert the code into the end of the predecessor block.
733 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
734 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
736 // Adjust the type back to match the PHI. Note that we can't use
737 // InsertPt here because the SCEVExpander may have inserted its
738 // instructions after that point, in its efforts to avoid inserting
739 // redundant expressions.
740 if (isa<PointerType>(PN->getType())) {
741 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
747 // Replace the use of the operand Value with the new Phi we just created.
748 PN->setIncomingValue(i, Code);
753 // PHI node might have become a constant value after SplitCriticalEdge.
754 DeadInsts.push_back(Inst);
756 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
760 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
761 /// mode, and does not need to be put in a register first.
762 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
763 const TargetLowering *TLI, bool HasBaseReg) {
764 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
765 int64_t VC = SC->getValue()->getSExtValue();
767 TargetLowering::AddrMode AM;
769 AM.HasBaseReg = HasBaseReg;
770 return TLI->isLegalAddressingMode(AM, UseTy);
772 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
773 return (VC > -(1 << 16) && VC < (1 << 16)-1);
777 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
778 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
779 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
780 Constant *Op0 = CE->getOperand(0);
781 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
782 TargetLowering::AddrMode AM;
784 AM.HasBaseReg = HasBaseReg;
785 return TLI->isLegalAddressingMode(AM, UseTy);
791 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
792 /// loop varying to the Imm operand.
793 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
794 Loop *L, ScalarEvolution *SE) {
795 if (Val->isLoopInvariant(L)) return; // Nothing to do.
797 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
798 std::vector<SCEVHandle> NewOps;
799 NewOps.reserve(SAE->getNumOperands());
801 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
802 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
803 // If this is a loop-variant expression, it must stay in the immediate
804 // field of the expression.
805 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
807 NewOps.push_back(SAE->getOperand(i));
811 Val = SE->getIntegerSCEV(0, Val->getType());
813 Val = SE->getAddExpr(NewOps);
814 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
815 // Try to pull immediates out of the start value of nested addrec's.
816 SCEVHandle Start = SARE->getStart();
817 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
819 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
821 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
823 // Otherwise, all of Val is variant, move the whole thing over.
824 Imm = SE->getAddExpr(Imm, Val);
825 Val = SE->getIntegerSCEV(0, Val->getType());
830 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
831 /// that can fit into the immediate field of instructions in the target.
832 /// Accumulate these immediate values into the Imm value.
833 static void MoveImmediateValues(const TargetLowering *TLI,
835 SCEVHandle &Val, SCEVHandle &Imm,
836 bool isAddress, Loop *L,
837 ScalarEvolution *SE) {
838 const Type *UseTy = User->getType();
839 if (StoreInst *SI = dyn_cast<StoreInst>(User))
840 UseTy = SI->getOperand(0)->getType();
842 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
843 std::vector<SCEVHandle> NewOps;
844 NewOps.reserve(SAE->getNumOperands());
846 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
847 SCEVHandle NewOp = SAE->getOperand(i);
848 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
850 if (!NewOp->isLoopInvariant(L)) {
851 // If this is a loop-variant expression, it must stay in the immediate
852 // field of the expression.
853 Imm = SE->getAddExpr(Imm, NewOp);
855 NewOps.push_back(NewOp);
860 Val = SE->getIntegerSCEV(0, Val->getType());
862 Val = SE->getAddExpr(NewOps);
864 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
865 // Try to pull immediates out of the start value of nested addrec's.
866 SCEVHandle Start = SARE->getStart();
867 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
869 if (Start != SARE->getStart()) {
870 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
872 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
875 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
876 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
877 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
878 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
880 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
881 SCEVHandle NewOp = SME->getOperand(1);
882 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
884 // If we extracted something out of the subexpressions, see if we can
886 if (NewOp != SME->getOperand(1)) {
887 // Scale SubImm up by "8". If the result is a target constant, we are
889 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
890 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
891 // Accumulate the immediate.
892 Imm = SE->getAddExpr(Imm, SubImm);
894 // Update what is left of 'Val'.
895 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
902 // Loop-variant expressions must stay in the immediate field of the
904 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
905 !Val->isLoopInvariant(L)) {
906 Imm = SE->getAddExpr(Imm, Val);
907 Val = SE->getIntegerSCEV(0, Val->getType());
911 // Otherwise, no immediates to move.
915 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
916 /// added together. This is used to reassociate common addition subexprs
917 /// together for maximal sharing when rewriting bases.
918 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
920 ScalarEvolution *SE) {
921 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
922 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
923 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
924 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
925 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
926 if (SARE->getOperand(0) == Zero) {
927 SubExprs.push_back(Expr);
929 // Compute the addrec with zero as its base.
930 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
931 Ops[0] = Zero; // Start with zero base.
932 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
935 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
937 } else if (!Expr->isZero()) {
939 SubExprs.push_back(Expr);
943 // This is logically local to the following function, but C++ says we have
944 // to make it file scope.
945 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
947 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
948 /// the Uses, removing any common subexpressions, except that if all such
949 /// subexpressions can be folded into an addressing mode for all uses inside
950 /// the loop (this case is referred to as "free" in comments herein) we do
951 /// not remove anything. This looks for things like (a+b+c) and
952 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
953 /// is *removed* from the Bases and returned.
955 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
956 ScalarEvolution *SE, Loop *L,
957 const TargetLowering *TLI) {
958 unsigned NumUses = Uses.size();
960 // Only one use? This is a very common case, so we handle it specially and
962 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
963 SCEVHandle Result = Zero;
964 SCEVHandle FreeResult = Zero;
966 // If the use is inside the loop, use its base, regardless of what it is:
967 // it is clearly shared across all the IV's. If the use is outside the loop
968 // (which means after it) we don't want to factor anything *into* the loop,
969 // so just use 0 as the base.
970 if (L->contains(Uses[0].Inst->getParent()))
971 std::swap(Result, Uses[0].Base);
975 // To find common subexpressions, count how many of Uses use each expression.
976 // If any subexpressions are used Uses.size() times, they are common.
977 // Also track whether all uses of each expression can be moved into an
978 // an addressing mode "for free"; such expressions are left within the loop.
979 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
980 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
982 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
983 // order we see them.
984 std::vector<SCEVHandle> UniqueSubExprs;
986 std::vector<SCEVHandle> SubExprs;
987 unsigned NumUsesInsideLoop = 0;
988 for (unsigned i = 0; i != NumUses; ++i) {
989 // If the user is outside the loop, just ignore it for base computation.
990 // Since the user is outside the loop, it must be *after* the loop (if it
991 // were before, it could not be based on the loop IV). We don't want users
992 // after the loop to affect base computation of values *inside* the loop,
993 // because we can always add their offsets to the result IV after the loop
994 // is done, ensuring we get good code inside the loop.
995 if (!L->contains(Uses[i].Inst->getParent()))
999 // If the base is zero (which is common), return zero now, there are no
1000 // CSEs we can find.
1001 if (Uses[i].Base == Zero) return Zero;
1003 // If this use is as an address we may be able to put CSEs in the addressing
1004 // mode rather than hoisting them.
1005 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1006 // We may need the UseTy below, but only when isAddrUse, so compute it
1007 // only in that case.
1008 const Type *UseTy = 0;
1010 UseTy = Uses[i].Inst->getType();
1011 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1012 UseTy = SI->getOperand(0)->getType();
1015 // Split the expression into subexprs.
1016 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1017 // Add one to SubExpressionUseData.Count for each subexpr present, and
1018 // if the subexpr is not a valid immediate within an addressing mode use,
1019 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1020 // hoist these out of the loop (if they are common to all uses).
1021 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1022 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1023 UniqueSubExprs.push_back(SubExprs[j]);
1024 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1025 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1030 // Now that we know how many times each is used, build Result. Iterate over
1031 // UniqueSubexprs so that we have a stable ordering.
1032 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1033 std::map<SCEVHandle, SubExprUseData>::iterator I =
1034 SubExpressionUseData.find(UniqueSubExprs[i]);
1035 assert(I != SubExpressionUseData.end() && "Entry not found?");
1036 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1037 if (I->second.notAllUsesAreFree)
1038 Result = SE->getAddExpr(Result, I->first);
1040 FreeResult = SE->getAddExpr(FreeResult, I->first);
1042 // Remove non-cse's from SubExpressionUseData.
1043 SubExpressionUseData.erase(I);
1046 if (FreeResult != Zero) {
1047 // We have some subexpressions that can be subsumed into addressing
1048 // modes in every use inside the loop. However, it's possible that
1049 // there are so many of them that the combined FreeResult cannot
1050 // be subsumed, or that the target cannot handle both a FreeResult
1051 // and a Result in the same instruction (for example because it would
1052 // require too many registers). Check this.
1053 for (unsigned i=0; i<NumUses; ++i) {
1054 if (!L->contains(Uses[i].Inst->getParent()))
1056 // We know this is an addressing mode use; if there are any uses that
1057 // are not, FreeResult would be Zero.
1058 const Type *UseTy = Uses[i].Inst->getType();
1059 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1060 UseTy = SI->getOperand(0)->getType();
1061 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1062 // FIXME: could split up FreeResult into pieces here, some hoisted
1063 // and some not. Doesn't seem worth it for now.
1064 Result = SE->getAddExpr(Result, FreeResult);
1071 // If we found no CSE's, return now.
1072 if (Result == Zero) return Result;
1074 // If we still have a FreeResult, remove its subexpressions from
1075 // SubExpressionUseData. This means they will remain in the use Bases.
1076 if (FreeResult != Zero) {
1077 SeparateSubExprs(SubExprs, FreeResult, SE);
1078 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1079 std::map<SCEVHandle, SubExprUseData>::iterator I =
1080 SubExpressionUseData.find(SubExprs[j]);
1081 SubExpressionUseData.erase(I);
1086 // Otherwise, remove all of the CSE's we found from each of the base values.
1087 for (unsigned i = 0; i != NumUses; ++i) {
1088 // Uses outside the loop don't necessarily include the common base, but
1089 // the final IV value coming into those uses does. Instead of trying to
1090 // remove the pieces of the common base, which might not be there,
1091 // subtract off the base to compensate for this.
1092 if (!L->contains(Uses[i].Inst->getParent())) {
1093 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1097 // Split the expression into subexprs.
1098 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1100 // Remove any common subexpressions.
1101 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1102 if (SubExpressionUseData.count(SubExprs[j])) {
1103 SubExprs.erase(SubExprs.begin()+j);
1107 // Finally, add the non-shared expressions together.
1108 if (SubExprs.empty())
1109 Uses[i].Base = Zero;
1111 Uses[i].Base = SE->getAddExpr(SubExprs);
1118 /// ValidStride - Check whether the given Scale is valid for all loads and
1119 /// stores in UsersToProcess.
1121 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1123 const std::vector<BasedUser>& UsersToProcess) {
1127 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1128 // If this is a load or other access, pass the type of the access in.
1129 const Type *AccessTy = Type::VoidTy;
1130 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1131 AccessTy = SI->getOperand(0)->getType();
1132 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1133 AccessTy = LI->getType();
1134 else if (isa<PHINode>(UsersToProcess[i].Inst))
1137 TargetLowering::AddrMode AM;
1138 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1139 AM.BaseOffs = SC->getValue()->getSExtValue();
1140 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1143 // If load[imm+r*scale] is illegal, bail out.
1144 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1150 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1152 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1156 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1158 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1159 !(isa<PointerType>(Ty2) &&
1160 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1161 !(isa<PointerType>(Ty1) &&
1162 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1165 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1166 /// of a previous stride and it is a legal value for the target addressing
1167 /// mode scale component and optional base reg. This allows the users of
1168 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1169 /// reuse is possible.
1170 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1171 bool AllUsesAreAddresses,
1172 const SCEVHandle &Stride,
1173 IVExpr &IV, const Type *Ty,
1174 const std::vector<BasedUser>& UsersToProcess) {
1175 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1176 int64_t SInt = SC->getValue()->getSExtValue();
1177 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1179 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1180 IVsByStride.find(StrideOrder[NewStride]);
1181 if (SI == IVsByStride.end())
1183 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1184 if (SI->first != Stride &&
1185 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1187 int64_t Scale = SInt / SSInt;
1188 // Check that this stride is valid for all the types used for loads and
1189 // stores; if it can be used for some and not others, we might as well use
1190 // the original stride everywhere, since we have to create the IV for it
1191 // anyway. If the scale is 1, then we don't need to worry about folding
1194 (AllUsesAreAddresses &&
1195 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1196 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1197 IE = SI->second.IVs.end(); II != IE; ++II)
1198 // FIXME: Only handle base == 0 for now.
1199 // Only reuse previous IV if it would not require a type conversion.
1200 if (II->Base->isZero() &&
1201 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1210 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1211 /// returns true if Val's isUseOfPostIncrementedValue is true.
1212 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1213 return Val.isUseOfPostIncrementedValue;
1216 /// isNonConstantNegative - Return true if the specified scev is negated, but
1218 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1219 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1220 if (!Mul) return false;
1222 // If there is a constant factor, it will be first.
1223 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1224 if (!SC) return false;
1226 // Return true if the value is negative, this matches things like (-42 * V).
1227 return SC->getValue()->getValue().isNegative();
1230 // CollectIVUsers - Transform our list of users and offsets to a bit more
1231 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1232 // of the strided accesses, as well as the old information from Uses. We
1233 // progressively move information from the Base field to the Imm field, until
1234 // we eventually have the full access expression to rewrite the use.
1235 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1236 IVUsersOfOneStride &Uses,
1238 bool &AllUsesAreAddresses,
1239 std::vector<BasedUser> &UsersToProcess) {
1240 UsersToProcess.reserve(Uses.Users.size());
1241 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1242 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1244 // Move any loop variant operands from the offset field to the immediate
1245 // field of the use, so that we don't try to use something before it is
1247 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1248 UsersToProcess.back().Imm, L, SE);
1249 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1250 "Base value is not loop invariant!");
1253 // We now have a whole bunch of uses of like-strided induction variables, but
1254 // they might all have different bases. We want to emit one PHI node for this
1255 // stride which we fold as many common expressions (between the IVs) into as
1256 // possible. Start by identifying the common expressions in the base values
1257 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1258 // "A+B"), emit it to the preheader, then remove the expression from the
1259 // UsersToProcess base values.
1260 SCEVHandle CommonExprs =
1261 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1263 // Next, figure out what we can represent in the immediate fields of
1264 // instructions. If we can represent anything there, move it to the imm
1265 // fields of the BasedUsers. We do this so that it increases the commonality
1266 // of the remaining uses.
1267 unsigned NumPHI = 0;
1268 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1269 // If the user is not in the current loop, this means it is using the exit
1270 // value of the IV. Do not put anything in the base, make sure it's all in
1271 // the immediate field to allow as much factoring as possible.
1272 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1273 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1274 UsersToProcess[i].Base);
1275 UsersToProcess[i].Base =
1276 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1279 // Addressing modes can be folded into loads and stores. Be careful that
1280 // the store is through the expression, not of the expression though.
1282 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1283 UsersToProcess[i].OperandValToReplace);
1284 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1289 // If this use isn't an address, then not all uses are addresses.
1290 if (!isAddress && !isPHI)
1291 AllUsesAreAddresses = false;
1293 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1294 UsersToProcess[i].Imm, isAddress, L, SE);
1298 // If one of the use if a PHI node and all other uses are addresses, still
1299 // allow iv reuse. Essentially we are trading one constant multiplication
1300 // for one fewer iv.
1302 AllUsesAreAddresses = false;
1307 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1308 /// stride of IV. All of the users may have different starting values, and this
1309 /// may not be the only stride (we know it is if isOnlyStride is true).
1310 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1311 IVUsersOfOneStride &Uses,
1313 bool isOnlyStride) {
1314 // If all the users are moved to another stride, then there is nothing to do.
1315 if (Uses.Users.empty())
1318 // Keep track if every use in UsersToProcess is an address. If they all are,
1319 // we may be able to rewrite the entire collection of them in terms of a
1320 // smaller-stride IV.
1321 bool AllUsesAreAddresses = true;
1323 // Transform our list of users and offsets to a bit more complex table. In
1324 // this new vector, each 'BasedUser' contains 'Base' the base of the
1325 // strided accessas well as the old information from Uses. We progressively
1326 // move information from the Base field to the Imm field, until we eventually
1327 // have the full access expression to rewrite the use.
1328 std::vector<BasedUser> UsersToProcess;
1329 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1332 // If we managed to find some expressions in common, we'll need to carry
1333 // their value in a register and add it in for each use. This will take up
1334 // a register operand, which potentially restricts what stride values are
1336 bool HaveCommonExprs = !CommonExprs->isZero();
1338 // If all uses are addresses, check if it is possible to reuse an IV with a
1339 // stride that is a factor of this stride. And that the multiple is a number
1340 // that can be encoded in the scale field of the target addressing mode. And
1341 // that we will have a valid instruction after this substition, including the
1342 // immediate field, if any.
1343 PHINode *NewPHI = NULL;
1345 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1346 SE->getIntegerSCEV(0, Type::Int32Ty),
1348 unsigned RewriteFactor = 0;
1349 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1350 Stride, ReuseIV, CommonExprs->getType(),
1352 if (RewriteFactor != 0) {
1353 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1354 << " and BASE " << *ReuseIV.Base << " :\n";
1355 NewPHI = ReuseIV.PHI;
1356 IncV = ReuseIV.IncV;
1359 const Type *ReplacedTy = CommonExprs->getType();
1361 // Now that we know what we need to do, insert the PHI node itself.
1363 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1364 << *Stride << " and BASE " << *CommonExprs << ": ";
1366 SCEVExpander Rewriter(*SE, *LI);
1367 SCEVExpander PreheaderRewriter(*SE, *LI);
1369 BasicBlock *Preheader = L->getLoopPreheader();
1370 Instruction *PreInsertPt = Preheader->getTerminator();
1371 Instruction *PhiInsertBefore = L->getHeader()->begin();
1373 BasicBlock *LatchBlock = L->getLoopLatch();
1376 // Emit the initial base value into the loop preheader.
1378 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1380 if (RewriteFactor == 0) {
1381 // Create a new Phi for this base, and stick it in the loop header.
1382 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1385 // Add common base to the new Phi node.
1386 NewPHI->addIncoming(CommonBaseV, Preheader);
1388 // If the stride is negative, insert a sub instead of an add for the
1390 bool isNegative = isNonConstantNegative(Stride);
1391 SCEVHandle IncAmount = Stride;
1393 IncAmount = SE->getNegativeSCEV(Stride);
1395 // Insert the stride into the preheader.
1396 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1397 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1399 // Emit the increment of the base value before the terminator of the loop
1400 // latch block, and add it to the Phi node.
1401 SCEVHandle IncExp = SE->getUnknown(StrideV);
1403 IncExp = SE->getNegativeSCEV(IncExp);
1404 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1406 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1407 IncV->setName(NewPHI->getName()+".inc");
1408 NewPHI->addIncoming(IncV, LatchBlock);
1410 // Remember this in case a later stride is multiple of this.
1411 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1413 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1415 Constant *C = dyn_cast<Constant>(CommonBaseV);
1417 (!C->isNullValue() &&
1418 !fitsInAddressMode(SE->getUnknown(CommonBaseV), ReplacedTy,
1420 // We want the common base emitted into the preheader! This is just
1421 // using cast as a copy so BitCast (no-op cast) is appropriate
1422 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1423 "commonbase", PreInsertPt);
1427 // We want to emit code for users inside the loop first. To do this, we
1428 // rearrange BasedUser so that the entries at the end have
1429 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1430 // vector (so we handle them first).
1431 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1432 PartitionByIsUseOfPostIncrementedValue);
1434 // Sort this by base, so that things with the same base are handled
1435 // together. By partitioning first and stable-sorting later, we are
1436 // guaranteed that within each base we will pop off users from within the
1437 // loop before users outside of the loop with a particular base.
1439 // We would like to use stable_sort here, but we can't. The problem is that
1440 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1441 // we don't have anything to do a '<' comparison on. Because we think the
1442 // number of uses is small, do a horrible bubble sort which just relies on
1444 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1445 // Get a base value.
1446 SCEVHandle Base = UsersToProcess[i].Base;
1448 // Compact everything with this base to be consequtive with this one.
1449 for (unsigned j = i+1; j != e; ++j) {
1450 if (UsersToProcess[j].Base == Base) {
1451 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1457 // Process all the users now. This outer loop handles all bases, the inner
1458 // loop handles all users of a particular base.
1459 while (!UsersToProcess.empty()) {
1460 SCEVHandle Base = UsersToProcess.back().Base;
1462 // Emit the code for Base into the preheader.
1463 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1465 DOUT << " INSERTING code for BASE = " << *Base << ":";
1466 if (BaseV->hasName())
1467 DOUT << " Result value name = %" << BaseV->getNameStr();
1470 // If BaseV is a constant other than 0, make sure that it gets inserted into
1471 // the preheader, instead of being forward substituted into the uses. We do
1472 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1474 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1475 if (!C->isNullValue() && !fitsInAddressMode(Base, ReplacedTy,
1477 // We want this constant emitted into the preheader! This is just
1478 // using cast as a copy so BitCast (no-op cast) is appropriate
1479 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1484 // Emit the code to add the immediate offset to the Phi value, just before
1485 // the instructions that we identified as using this stride and base.
1487 // FIXME: Use emitted users to emit other users.
1488 BasedUser &User = UsersToProcess.back();
1490 // If this instruction wants to use the post-incremented value, move it
1491 // after the post-inc and use its value instead of the PHI.
1492 Value *RewriteOp = NewPHI;
1493 if (User.isUseOfPostIncrementedValue) {
1496 // If this user is in the loop, make sure it is the last thing in the
1497 // loop to ensure it is dominated by the increment.
1498 if (L->contains(User.Inst->getParent()))
1499 User.Inst->moveBefore(LatchBlock->getTerminator());
1501 if (RewriteOp->getType() != ReplacedTy) {
1502 Instruction::CastOps opcode = Instruction::Trunc;
1503 if (ReplacedTy->getPrimitiveSizeInBits() ==
1504 RewriteOp->getType()->getPrimitiveSizeInBits())
1505 opcode = Instruction::BitCast;
1506 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1509 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1511 // If we had to insert new instrutions for RewriteOp, we have to
1512 // consider that they may not have been able to end up immediately
1513 // next to RewriteOp, because non-PHI instructions may never precede
1514 // PHI instructions in a block. In this case, remember where the last
1515 // instruction was inserted so that if we're replacing a different
1516 // PHI node, we can use the later point to expand the final
1518 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1519 if (RewriteOp == NewPHI) NewBasePt = 0;
1521 // Clear the SCEVExpander's expression map so that we are guaranteed
1522 // to have the code emitted where we expect it.
1525 // If we are reusing the iv, then it must be multiplied by a constant
1526 // factor take advantage of addressing mode scale component.
1527 if (RewriteFactor != 0) {
1528 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1529 RewriteExpr->getType()),
1532 // The common base is emitted in the loop preheader. But since we
1533 // are reusing an IV, it has not been used to initialize the PHI node.
1534 // Add it to the expression used to rewrite the uses.
1535 if (!isa<ConstantInt>(CommonBaseV) ||
1536 !cast<ConstantInt>(CommonBaseV)->isZero())
1537 RewriteExpr = SE->getAddExpr(RewriteExpr,
1538 SE->getUnknown(CommonBaseV));
1541 // Now that we know what we need to do, insert code before User for the
1542 // immediate and any loop-variant expressions.
1543 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1544 // Add BaseV to the PHI value if needed.
1545 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1547 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1551 // Mark old value we replaced as possibly dead, so that it is eliminated
1552 // if we just replaced the last use of that value.
1553 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1555 UsersToProcess.pop_back();
1558 // If there are any more users to process with the same base, process them
1559 // now. We sorted by base above, so we just have to check the last elt.
1560 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1561 // TODO: Next, find out which base index is the most common, pull it out.
1564 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1565 // different starting values, into different PHIs.
1568 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1569 /// set the IV user and stride information and return true, otherwise return
1571 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1572 const SCEVHandle *&CondStride) {
1573 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1575 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1576 IVUsesByStride.find(StrideOrder[Stride]);
1577 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1579 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1580 E = SI->second.Users.end(); UI != E; ++UI)
1581 if (UI->User == Cond) {
1582 // NOTE: we could handle setcc instructions with multiple uses here, but
1583 // InstCombine does it as well for simple uses, it's not clear that it
1584 // occurs enough in real life to handle.
1586 CondStride = &SI->first;
1594 // Constant strides come first which in turns are sorted by their absolute
1595 // values. If absolute values are the same, then positive strides comes first.
1597 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1598 struct StrideCompare {
1599 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1600 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1601 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1603 int64_t LV = LHSC->getValue()->getSExtValue();
1604 int64_t RV = RHSC->getValue()->getSExtValue();
1605 uint64_t ALV = (LV < 0) ? -LV : LV;
1606 uint64_t ARV = (RV < 0) ? -RV : RV;
1612 return (LHSC && !RHSC);
1617 /// ChangeCompareStride - If a loop termination compare instruction is the
1618 /// only use of its stride, and the compaison is against a constant value,
1619 /// try eliminate the stride by moving the compare instruction to another
1620 /// stride and change its constant operand accordingly. e.g.
1626 /// if (v2 < 10) goto loop
1631 /// if (v1 < 30) goto loop
1632 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1633 IVStrideUse* &CondUse,
1634 const SCEVHandle* &CondStride) {
1635 if (StrideOrder.size() < 2 ||
1636 IVUsesByStride[*CondStride].Users.size() != 1)
1638 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1639 if (!SC) return Cond;
1640 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1641 if (!C) return Cond;
1643 ICmpInst::Predicate Predicate = Cond->getPredicate();
1644 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1645 int64_t CmpVal = C->getValue().getSExtValue();
1646 unsigned BitWidth = C->getValue().getBitWidth();
1647 uint64_t SignBit = 1ULL << (BitWidth-1);
1648 const Type *CmpTy = C->getType();
1649 const Type *NewCmpTy = NULL;
1650 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1651 unsigned NewTyBits = 0;
1652 int64_t NewCmpVal = CmpVal;
1653 SCEVHandle *NewStride = NULL;
1654 Value *NewIncV = NULL;
1657 // Check stride constant and the comparision constant signs to detect
1659 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1662 // Look for a suitable stride / iv as replacement.
1663 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1664 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1665 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1666 IVUsesByStride.find(StrideOrder[i]);
1667 if (!isa<SCEVConstant>(SI->first))
1669 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1670 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1673 Scale = SSInt / CmpSSInt;
1674 NewCmpVal = CmpVal * Scale;
1675 APInt Mul = APInt(BitWidth, NewCmpVal);
1676 // Check for overflow.
1677 if (Mul.getSExtValue() != NewCmpVal) {
1682 // Watch out for overflow.
1683 if (ICmpInst::isSignedPredicate(Predicate) &&
1684 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1687 if (NewCmpVal != CmpVal) {
1688 // Pick the best iv to use trying to avoid a cast.
1690 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1691 E = SI->second.Users.end(); UI != E; ++UI) {
1692 NewIncV = UI->OperandValToReplace;
1693 if (NewIncV->getType() == CmpTy)
1701 NewCmpTy = NewIncV->getType();
1702 NewTyBits = isa<PointerType>(NewCmpTy)
1703 ? UIntPtrTy->getPrimitiveSizeInBits()
1704 : NewCmpTy->getPrimitiveSizeInBits();
1705 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1706 // Check if it is possible to rewrite it using
1707 // an iv / stride of a smaller integer type.
1708 bool TruncOk = false;
1709 if (NewCmpTy->isInteger()) {
1710 unsigned Bits = NewTyBits;
1711 if (ICmpInst::isSignedPredicate(Predicate))
1713 uint64_t Mask = (1ULL << Bits) - 1;
1714 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1723 // Don't rewrite if use offset is non-constant and the new type is
1724 // of a different type.
1725 // FIXME: too conservative?
1726 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1731 bool AllUsesAreAddresses = true;
1732 std::vector<BasedUser> UsersToProcess;
1733 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1734 AllUsesAreAddresses,
1736 // Avoid rewriting the compare instruction with an iv of new stride
1737 // if it's likely the new stride uses will be rewritten using the
1738 if (AllUsesAreAddresses &&
1739 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1744 // If scale is negative, use swapped predicate unless it's testing
1746 if (Scale < 0 && !Cond->isEquality())
1747 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1749 NewStride = &StrideOrder[i];
1754 // Forgo this transformation if it the increment happens to be
1755 // unfortunately positioned after the condition, and the condition
1756 // has multiple uses which prevent it from being moved immediately
1757 // before the branch. See
1758 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1759 // for an example of this situation.
1760 if (!Cond->hasOneUse()) {
1761 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1767 if (NewCmpVal != CmpVal) {
1768 // Create a new compare instruction using new stride / iv.
1769 ICmpInst *OldCond = Cond;
1771 if (!isa<PointerType>(NewCmpTy))
1772 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1774 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1775 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1777 // Insert new compare instruction.
1778 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1779 L->getHeader()->getName() + ".termcond",
1782 // Remove the old compare instruction. The old indvar is probably dead too.
1783 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
1784 SE->deleteValueFromRecords(OldCond);
1785 OldCond->replaceAllUsesWith(Cond);
1786 OldCond->eraseFromParent();
1788 IVUsesByStride[*CondStride].Users.pop_back();
1789 SCEVHandle NewOffset = TyBits == NewTyBits
1790 ? SE->getMulExpr(CondUse->Offset,
1791 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1792 : SE->getConstant(ConstantInt::get(NewCmpTy,
1793 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1794 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1795 CondUse = &IVUsesByStride[*NewStride].Users.back();
1796 CondStride = NewStride;
1803 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1804 /// an smax computation.
1806 /// This is a narrow solution to a specific, but acute, problem. For loops
1812 /// } while (++i < n);
1814 /// where the comparison is signed, the trip count isn't just 'n', because
1815 /// 'n' could be negative. And unfortunately this can come up even for loops
1816 /// where the user didn't use a C do-while loop. For example, seemingly
1817 /// well-behaved top-test loops will commonly be lowered like this:
1823 /// } while (++i < n);
1826 /// and then it's possible for subsequent optimization to obscure the if
1827 /// test in such a way that indvars can't find it.
1829 /// When indvars can't find the if test in loops like this, it creates a
1830 /// signed-max expression, which allows it to give the loop a canonical
1831 /// induction variable:
1834 /// smax = n < 1 ? 1 : n;
1837 /// } while (++i != smax);
1839 /// Canonical induction variables are necessary because the loop passes
1840 /// are designed around them. The most obvious example of this is the
1841 /// LoopInfo analysis, which doesn't remember trip count values. It
1842 /// expects to be able to rediscover the trip count each time it is
1843 /// needed, and it does this using a simple analyis that only succeeds if
1844 /// the loop has a canonical induction variable.
1846 /// However, when it comes time to generate code, the maximum operation
1847 /// can be quite costly, especially if it's inside of an outer loop.
1849 /// This function solves this problem by detecting this type of loop and
1850 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1851 /// the instructions for the maximum computation.
1853 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1854 IVStrideUse* &CondUse) {
1855 // Check that the loop matches the pattern we're looking for.
1856 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1857 Cond->getPredicate() != CmpInst::ICMP_NE)
1860 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1861 if (!Sel || !Sel->hasOneUse()) return Cond;
1863 SCEVHandle IterationCount = SE->getIterationCount(L);
1864 if (isa<SCEVCouldNotCompute>(IterationCount))
1866 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1868 // Adjust for an annoying getIterationCount quirk.
1869 IterationCount = SE->getAddExpr(IterationCount, One);
1871 // Check for a max calculation that matches the pattern.
1872 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1873 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1875 SCEVHandle SMaxLHS = SMax->getOperand(0);
1876 SCEVHandle SMaxRHS = SMax->getOperand(1);
1877 if (!SMaxLHS || SMaxLHS != One) return Cond;
1879 // Check the relevant induction variable for conformance to
1881 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1882 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1883 if (!AR || !AR->isAffine() ||
1884 AR->getStart() != One ||
1885 AR->getStepRecurrence(*SE) != One)
1888 // Check the right operand of the select, and remember it, as it will
1889 // be used in the new comparison instruction.
1891 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1892 NewRHS = Sel->getOperand(1);
1893 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1894 NewRHS = Sel->getOperand(2);
1895 if (!NewRHS) return Cond;
1897 // Ok, everything looks ok to change the condition into an SLT or SGE and
1898 // delete the max calculation.
1900 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
1903 Cond->getOperand(0), NewRHS, "scmp", Cond);
1905 // Delete the max calculation instructions.
1906 SE->deleteValueFromRecords(Cond);
1907 Cond->replaceAllUsesWith(NewCond);
1908 Cond->eraseFromParent();
1909 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1910 SE->deleteValueFromRecords(Sel);
1911 Sel->eraseFromParent();
1912 if (Cmp->use_empty()) {
1913 SE->deleteValueFromRecords(Cmp);
1914 Cmp->eraseFromParent();
1916 CondUse->User = NewCond;
1920 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1921 /// inside the loop then try to eliminate the cast opeation.
1922 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1924 SCEVHandle IterationCount = SE->getIterationCount(L);
1925 if (isa<SCEVCouldNotCompute>(IterationCount))
1928 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1930 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1931 IVUsesByStride.find(StrideOrder[Stride]);
1932 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1933 if (!isa<SCEVConstant>(SI->first))
1936 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1937 E = SI->second.Users.end(); UI != E; /* empty */) {
1938 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1940 Instruction *ShadowUse = CandidateUI->User;
1941 const Type *DestTy = NULL;
1943 /* If shadow use is a int->float cast then insert a second IV
1944 to eliminate this cast.
1946 for (unsigned i = 0; i < n; ++i)
1952 for (unsigned i = 0; i < n; ++i, ++d)
1955 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1956 DestTy = UCast->getDestTy();
1957 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1958 DestTy = SCast->getDestTy();
1959 if (!DestTy) continue;
1962 /* If target does not support DestTy natively then do not apply
1963 this transformation. */
1964 MVT DVT = TLI->getValueType(DestTy);
1965 if (!TLI->isTypeLegal(DVT)) continue;
1968 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1970 if (PH->getNumIncomingValues() != 2) continue;
1972 const Type *SrcTy = PH->getType();
1973 int Mantissa = DestTy->getFPMantissaWidth();
1974 if (Mantissa == -1) continue;
1975 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1978 unsigned Entry, Latch;
1979 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1987 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1988 if (!Init) continue;
1989 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1991 BinaryOperator *Incr =
1992 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1993 if (!Incr) continue;
1994 if (Incr->getOpcode() != Instruction::Add
1995 && Incr->getOpcode() != Instruction::Sub)
1998 /* Initialize new IV, double d = 0.0 in above example. */
1999 ConstantInt *C = NULL;
2000 if (Incr->getOperand(0) == PH)
2001 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2002 else if (Incr->getOperand(1) == PH)
2003 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2009 /* Add new PHINode. */
2010 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2012 /* create new increment. '++d' in above example. */
2013 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2014 BinaryOperator *NewIncr =
2015 BinaryOperator::Create(Incr->getOpcode(),
2016 NewPH, CFP, "IV.S.next.", Incr);
2018 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2019 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2021 /* Remove cast operation */
2022 SE->deleteValueFromRecords(ShadowUse);
2023 ShadowUse->replaceAllUsesWith(NewPH);
2024 ShadowUse->eraseFromParent();
2025 SI->second.Users.erase(CandidateUI);
2032 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2033 // uses in the loop, look to see if we can eliminate some, in favor of using
2034 // common indvars for the different uses.
2035 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2036 // TODO: implement optzns here.
2038 OptimizeShadowIV(L);
2040 // Finally, get the terminating condition for the loop if possible. If we
2041 // can, we want to change it to use a post-incremented version of its
2042 // induction variable, to allow coalescing the live ranges for the IV into
2043 // one register value.
2044 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2045 BasicBlock *Preheader = L->getLoopPreheader();
2046 BasicBlock *LatchBlock =
2047 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2048 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2049 if (!TermBr || TermBr->isUnconditional() ||
2050 !isa<ICmpInst>(TermBr->getCondition()))
2052 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2054 // Search IVUsesByStride to find Cond's IVUse if there is one.
2055 IVStrideUse *CondUse = 0;
2056 const SCEVHandle *CondStride = 0;
2058 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2059 return; // setcc doesn't use the IV.
2061 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2062 // being unable to find a sufficient guard, for example), change the loop
2063 // comparison to use SLT instead of NE.
2064 Cond = OptimizeSMax(L, Cond, CondUse);
2066 // If possible, change stride and operands of the compare instruction to
2067 // eliminate one stride.
2068 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2070 // It's possible for the setcc instruction to be anywhere in the loop, and
2071 // possible for it to have multiple users. If it is not immediately before
2072 // the latch block branch, move it.
2073 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2074 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2075 Cond->moveBefore(TermBr);
2077 // Otherwise, clone the terminating condition and insert into the loopend.
2078 Cond = cast<ICmpInst>(Cond->clone());
2079 Cond->setName(L->getHeader()->getName() + ".termcond");
2080 LatchBlock->getInstList().insert(TermBr, Cond);
2082 // Clone the IVUse, as the old use still exists!
2083 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2084 CondUse->OperandValToReplace);
2085 CondUse = &IVUsesByStride[*CondStride].Users.back();
2089 // If we get to here, we know that we can transform the setcc instruction to
2090 // use the post-incremented version of the IV, allowing us to coalesce the
2091 // live ranges for the IV correctly.
2092 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2093 CondUse->isUseOfPostIncrementedValue = true;
2097 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2099 LI = &getAnalysis<LoopInfo>();
2100 DT = &getAnalysis<DominatorTree>();
2101 SE = &getAnalysis<ScalarEvolution>();
2102 TD = &getAnalysis<TargetData>();
2103 UIntPtrTy = TD->getIntPtrType();
2106 // Find all uses of induction variables in this loop, and catagorize
2107 // them by stride. Start by finding all of the PHI nodes in the header for
2108 // this loop. If they are induction variables, inspect their uses.
2109 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2110 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2111 AddUsersIfInteresting(I, L, Processed);
2113 if (!IVUsesByStride.empty()) {
2114 // Optimize induction variables. Some indvar uses can be transformed to use
2115 // strides that will be needed for other purposes. A common example of this
2116 // is the exit test for the loop, which can often be rewritten to use the
2117 // computation of some other indvar to decide when to terminate the loop.
2120 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2121 // doing computation in byte values, promote to 32-bit values if safe.
2123 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2124 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2125 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2126 // Need to be careful that IV's are all the same type. Only works for
2127 // intptr_t indvars.
2129 // If we only have one stride, we can more aggressively eliminate some
2131 bool HasOneStride = IVUsesByStride.size() == 1;
2134 DOUT << "\nLSR on ";
2138 // IVsByStride keeps IVs for one particular loop.
2139 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2141 // Sort the StrideOrder so we process larger strides first.
2142 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2144 // Note: this processes each stride/type pair individually. All users
2145 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2146 // Also, note that we iterate over IVUsesByStride indirectly by using
2147 // StrideOrder. This extra layer of indirection makes the ordering of
2148 // strides deterministic - not dependent on map order.
2149 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2150 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2151 IVUsesByStride.find(StrideOrder[Stride]);
2152 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2153 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2157 // We're done analyzing this loop; release all the state we built up for it.
2158 CastedPointers.clear();
2159 IVUsesByStride.clear();
2160 IVsByStride.clear();
2161 StrideOrder.clear();
2163 // Clean up after ourselves
2164 if (!DeadInsts.empty()) {
2165 DeleteTriviallyDeadInstructions();
2167 BasicBlock::iterator I = L->getHeader()->begin();
2168 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2169 // At this point, we know that we have killed one or more IV users.
2170 // It is worth checking to see if the cannonical indvar is also
2171 // dead, so that we can remove it as well.
2173 // We can remove a PHI if it is on a cycle in the def-use graph
2174 // where each node in the cycle has degree one, i.e. only one use,
2175 // and is an instruction with no side effects.
2177 // FIXME: this needs to eliminate an induction variable even if it's being
2178 // compared against some value to decide loop termination.
2179 if (!PN->hasOneUse())
2182 SmallPtrSet<PHINode *, 4> PHIs;
2183 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2184 J && J->hasOneUse() && !J->mayWriteToMemory();
2185 J = dyn_cast<Instruction>(*J->use_begin())) {
2186 // If we find the original PHI, we've discovered a cycle.
2188 // Break the cycle and mark the PHI for deletion.
2189 SE->deleteValueFromRecords(PN);
2190 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2191 DeadInsts.push_back(PN);
2195 // If we find a PHI more than once, we're on a cycle that
2196 // won't prove fruitful.
2197 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2201 DeleteTriviallyDeadInstructions();