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 /// GEPlist - A list of the GEP's that have been remembered in the SCEV
134 /// data structures. SCEV does not know to update these when the operands
135 /// of the GEP are changed, which means we cannot leave them live across
137 SmallVector<GetElementPtrInst *, 16> GEPlist;
139 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
140 /// of the casted version of each value. This is accessed by
141 /// getCastedVersionOf.
142 DenseMap<Value*, Value*> CastedPointers;
144 /// DeadInsts - Keep track of instructions we may have made dead, so that
145 /// we can remove them after we are done working.
146 SmallVector<Instruction*, 16> DeadInsts;
148 /// TLI - Keep a pointer of a TargetLowering to consult for determining
149 /// transformation profitability.
150 const TargetLowering *TLI;
153 static char ID; // Pass ID, replacement for typeid
154 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
155 LoopPass(&ID), TLI(tli) {
158 bool runOnLoop(Loop *L, LPPassManager &LPM);
160 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
161 // We split critical edges, so we change the CFG. However, we do update
162 // many analyses if they are around.
163 AU.addPreservedID(LoopSimplifyID);
164 AU.addPreserved<LoopInfo>();
165 AU.addPreserved<DominanceFrontier>();
166 AU.addPreserved<DominatorTree>();
168 AU.addRequiredID(LoopSimplifyID);
169 AU.addRequired<LoopInfo>();
170 AU.addRequired<DominatorTree>();
171 AU.addRequired<TargetData>();
172 AU.addRequired<ScalarEvolution>();
173 AU.addPreserved<ScalarEvolution>();
176 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
178 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
180 bool AddUsersIfInteresting(Instruction *I, Loop *L,
181 SmallPtrSet<Instruction*,16> &Processed);
182 SCEVHandle GetExpressionSCEV(Instruction *E);
183 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
184 IVStrideUse* &CondUse,
185 const SCEVHandle* &CondStride);
186 void OptimizeIndvars(Loop *L);
188 /// OptimizeShadowIV - If IV is used in a int-to-float cast
189 /// inside the loop then try to eliminate the cast opeation.
190 void OptimizeShadowIV(Loop *L);
192 /// OptimizeSMax - Rewrite the loop's terminating condition
193 /// if it uses an smax computation.
194 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
195 IVStrideUse* &CondUse);
197 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
198 const SCEVHandle *&CondStride);
199 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
200 SCEVHandle CheckForIVReuse(bool, bool, bool, const SCEVHandle&,
201 IVExpr&, const Type*,
202 const std::vector<BasedUser>& UsersToProcess);
203 bool ValidStride(bool, int64_t,
204 const std::vector<BasedUser>& UsersToProcess);
205 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
206 IVUsersOfOneStride &Uses,
208 bool &AllUsesAreAddresses,
209 bool &AllUsesAreOutsideLoop,
210 std::vector<BasedUser> &UsersToProcess);
211 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
212 IVUsersOfOneStride &Uses,
213 Loop *L, bool isOnlyStride);
214 void DeleteTriviallyDeadInstructions();
218 char LoopStrengthReduce::ID = 0;
219 static RegisterPass<LoopStrengthReduce>
220 X("loop-reduce", "Loop Strength Reduction");
222 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
223 return new LoopStrengthReduce(TLI);
226 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
227 /// assumes that the Value* V is of integer or pointer type only.
229 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
231 if (V->getType() == UIntPtrTy) return V;
232 if (Constant *CB = dyn_cast<Constant>(V))
233 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
235 Value *&New = CastedPointers[V];
238 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
239 DeadInsts.push_back(cast<Instruction>(New));
244 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
245 /// specified set are trivially dead, delete them and see if this makes any of
246 /// their operands subsequently dead.
247 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
248 if (DeadInsts.empty()) return;
250 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
251 // go. The code below never adds a non-dead instruction to the worklist, but
252 // callers may not be so careful.
253 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
255 // Drop duplicate instructions and those with uses.
256 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
257 Instruction *I = DeadInsts[i];
258 if (!I->use_empty()) DeadInsts[i] = 0;
259 while (i != e && DeadInsts[i+1] == I)
263 while (!DeadInsts.empty()) {
264 Instruction *I = DeadInsts.back();
265 DeadInsts.pop_back();
267 if (I == 0 || !isInstructionTriviallyDead(I))
270 SE->deleteValueFromRecords(I);
272 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
273 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
276 DeadInsts.push_back(U);
280 I->eraseFromParent();
286 /// GetExpressionSCEV - Compute and return the SCEV for the specified
288 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
289 // Pointer to pointer bitcast instructions return the same value as their
291 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
292 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
293 return SE->getSCEV(BCI);
294 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
299 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
300 // If this is a GEP that SE doesn't know about, compute it now and insert it.
301 // If this is not a GEP, or if we have already done this computation, just let
303 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
304 if (!GEP || SE->hasSCEV(GEP))
305 return SE->getSCEV(Exp);
307 // Analyze all of the subscripts of this getelementptr instruction, looking
308 // for uses that are determined by the trip count of the loop. First, skip
309 // all operands the are not dependent on the IV.
311 // Build up the base expression. Insert an LLVM cast of the pointer to
313 SCEVHandle GEPVal = SE->getUnknown(
314 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
316 gep_type_iterator GTI = gep_type_begin(GEP);
318 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
319 i != e; ++i, ++GTI) {
320 // If this is a use of a recurrence that we can analyze, and it comes before
321 // Op does in the GEP operand list, we will handle this when we process this
323 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
324 const StructLayout *SL = TD->getStructLayout(STy);
325 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
326 uint64_t Offset = SL->getElementOffset(Idx);
327 GEPVal = SE->getAddExpr(GEPVal,
328 SE->getIntegerSCEV(Offset, UIntPtrTy));
330 unsigned GEPOpiBits =
331 (*i)->getType()->getPrimitiveSizeInBits();
332 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
333 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
334 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
335 Instruction::BitCast));
336 Value *OpVal = getCastedVersionOf(opcode, *i);
337 SCEVHandle Idx = SE->getSCEV(OpVal);
339 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
341 Idx = SE->getMulExpr(Idx,
342 SE->getConstant(ConstantInt::get(UIntPtrTy,
344 GEPVal = SE->getAddExpr(GEPVal, Idx);
348 SE->setSCEV(GEP, GEPVal);
349 GEPlist.push_back(GEP);
353 /// getSCEVStartAndStride - Compute the start and stride of this expression,
354 /// returning false if the expression is not a start/stride pair, or true if it
355 /// is. The stride must be a loop invariant expression, but the start may be
356 /// a mix of loop invariant and loop variant expressions.
357 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
358 SCEVHandle &Start, SCEVHandle &Stride,
359 ScalarEvolution *SE) {
360 SCEVHandle TheAddRec = Start; // Initialize to zero.
362 // If the outer level is an AddExpr, the operands are all start values except
363 // for a nested AddRecExpr.
364 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
365 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
366 if (SCEVAddRecExpr *AddRec =
367 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
368 if (AddRec->getLoop() == L)
369 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
371 return false; // Nested IV of some sort?
373 Start = SE->getAddExpr(Start, AE->getOperand(i));
376 } else if (isa<SCEVAddRecExpr>(SH)) {
379 return false; // not analyzable.
382 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
383 if (!AddRec || AddRec->getLoop() != L) return false;
385 // FIXME: Generalize to non-affine IV's.
386 if (!AddRec->isAffine()) return false;
388 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
390 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
391 DOUT << "[" << L->getHeader()->getName()
392 << "] Variable stride: " << *AddRec << "\n";
394 Stride = AddRec->getOperand(1);
398 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
399 /// and now we need to decide whether the user should use the preinc or post-inc
400 /// value. If this user should use the post-inc version of the IV, return true.
402 /// Choosing wrong here can break dominance properties (if we choose to use the
403 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
404 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
405 /// should use the post-inc value).
406 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
407 Loop *L, DominatorTree *DT, Pass *P,
408 SmallVectorImpl<Instruction*> &DeadInsts){
409 // If the user is in the loop, use the preinc value.
410 if (L->contains(User->getParent())) return false;
412 BasicBlock *LatchBlock = L->getLoopLatch();
414 // Ok, the user is outside of the loop. If it is dominated by the latch
415 // block, use the post-inc value.
416 if (DT->dominates(LatchBlock, User->getParent()))
419 // There is one case we have to be careful of: PHI nodes. These little guys
420 // can live in blocks that do not dominate the latch block, but (since their
421 // uses occur in the predecessor block, not the block the PHI lives in) should
422 // still use the post-inc value. Check for this case now.
423 PHINode *PN = dyn_cast<PHINode>(User);
424 if (!PN) return false; // not a phi, not dominated by latch block.
426 // Look at all of the uses of IV by the PHI node. If any use corresponds to
427 // a block that is not dominated by the latch block, give up and use the
428 // preincremented value.
429 unsigned NumUses = 0;
430 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
431 if (PN->getIncomingValue(i) == IV) {
433 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
437 // Okay, all uses of IV by PN are in predecessor blocks that really are
438 // dominated by the latch block. Split the critical edges and use the
439 // post-incremented value.
440 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
441 if (PN->getIncomingValue(i) == IV) {
442 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
443 // Splitting the critical edge can reduce the number of entries in this
445 e = PN->getNumIncomingValues();
446 if (--NumUses == 0) break;
449 // PHI node might have become a constant value after SplitCriticalEdge.
450 DeadInsts.push_back(User);
455 /// isAddress - Returns true if the specified instruction is using the
456 /// specified value as an address.
457 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
458 bool isAddress = isa<LoadInst>(Inst);
459 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
460 if (SI->getOperand(1) == OperandVal)
462 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
463 // Addressing modes can also be folded into prefetches and a variety
465 switch (II->getIntrinsicID()) {
467 case Intrinsic::prefetch:
468 case Intrinsic::x86_sse2_loadu_dq:
469 case Intrinsic::x86_sse2_loadu_pd:
470 case Intrinsic::x86_sse_loadu_ps:
471 case Intrinsic::x86_sse_storeu_ps:
472 case Intrinsic::x86_sse2_storeu_pd:
473 case Intrinsic::x86_sse2_storeu_dq:
474 case Intrinsic::x86_sse2_storel_dq:
475 if (II->getOperand(1) == OperandVal)
483 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
484 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
485 /// return true. Otherwise, return false.
486 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
487 SmallPtrSet<Instruction*,16> &Processed) {
488 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
489 return false; // Void and FP expressions cannot be reduced.
490 if (!Processed.insert(I))
491 return true; // Instruction already handled.
493 // Get the symbolic expression for this instruction.
494 SCEVHandle ISE = GetExpressionSCEV(I);
495 if (isa<SCEVCouldNotCompute>(ISE)) return false;
497 // Get the start and stride for this expression.
498 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
499 SCEVHandle Stride = Start;
500 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
501 return false; // Non-reducible symbolic expression, bail out.
503 std::vector<Instruction *> IUsers;
504 // Collect all I uses now because IVUseShouldUsePostIncValue may
505 // invalidate use_iterator.
506 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
507 IUsers.push_back(cast<Instruction>(*UI));
509 for (unsigned iused_index = 0, iused_size = IUsers.size();
510 iused_index != iused_size; ++iused_index) {
512 Instruction *User = IUsers[iused_index];
514 // Do not infinitely recurse on PHI nodes.
515 if (isa<PHINode>(User) && Processed.count(User))
518 // Descend recursively, but not into PHI nodes outside the current loop.
519 // It's important to see the entire expression outside the loop to get
520 // choices that depend on addressing mode use right, although we won't
521 // consider references ouside the loop in all cases.
522 // If User is already in Processed, we don't want to recurse into it again,
523 // but do want to record a second reference in the same instruction.
524 bool AddUserToIVUsers = false;
525 if (LI->getLoopFor(User->getParent()) != L) {
526 if (isa<PHINode>(User) || Processed.count(User) ||
527 !AddUsersIfInteresting(User, L, Processed)) {
528 DOUT << "FOUND USER in other loop: " << *User
529 << " OF SCEV: " << *ISE << "\n";
530 AddUserToIVUsers = true;
532 } else if (Processed.count(User) ||
533 !AddUsersIfInteresting(User, L, Processed)) {
534 DOUT << "FOUND USER: " << *User
535 << " OF SCEV: " << *ISE << "\n";
536 AddUserToIVUsers = true;
539 if (AddUserToIVUsers) {
540 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
541 if (StrideUses.Users.empty()) // First occurrence of this stride?
542 StrideOrder.push_back(Stride);
544 // Okay, we found a user that we cannot reduce. Analyze the instruction
545 // and decide what to do with it. If we are a use inside of the loop, use
546 // the value before incrementation, otherwise use it after incrementation.
547 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
548 // The value used will be incremented by the stride more than we are
549 // expecting, so subtract this off.
550 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
551 StrideUses.addUser(NewStart, User, I);
552 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
553 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
555 StrideUses.addUser(Start, User, I);
563 /// BasedUser - For a particular base value, keep information about how we've
564 /// partitioned the expression so far.
566 /// SE - The current ScalarEvolution object.
569 /// Base - The Base value for the PHI node that needs to be inserted for
570 /// this use. As the use is processed, information gets moved from this
571 /// field to the Imm field (below). BasedUser values are sorted by this
575 /// Inst - The instruction using the induction variable.
578 /// OperandValToReplace - The operand value of Inst to replace with the
580 Value *OperandValToReplace;
582 /// Imm - The immediate value that should be added to the base immediately
583 /// before Inst, because it will be folded into the imm field of the
587 // isUseOfPostIncrementedValue - True if this should use the
588 // post-incremented version of this IV, not the preincremented version.
589 // This can only be set in special cases, such as the terminating setcc
590 // instruction for a loop and uses outside the loop that are dominated by
592 bool isUseOfPostIncrementedValue;
594 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
595 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
596 OperandValToReplace(IVSU.OperandValToReplace),
597 Imm(SE->getIntegerSCEV(0, Base->getType())),
598 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
600 // Once we rewrite the code to insert the new IVs we want, update the
601 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
603 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
604 Instruction *InsertPt,
605 SCEVExpander &Rewriter, Loop *L, Pass *P,
606 SmallVectorImpl<Instruction*> &DeadInsts);
608 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
609 SCEVExpander &Rewriter,
610 Instruction *IP, Loop *L);
615 void BasedUser::dump() const {
616 cerr << " Base=" << *Base;
617 cerr << " Imm=" << *Imm;
618 cerr << " Inst: " << *Inst;
621 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
622 SCEVExpander &Rewriter,
623 Instruction *IP, Loop *L) {
624 // Figure out where we *really* want to insert this code. In particular, if
625 // the user is inside of a loop that is nested inside of L, we really don't
626 // want to insert this expression before the user, we'd rather pull it out as
627 // many loops as possible.
628 LoopInfo &LI = Rewriter.getLoopInfo();
629 Instruction *BaseInsertPt = IP;
631 // Figure out the most-nested loop that IP is in.
632 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
634 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
635 // the preheader of the outer-most loop where NewBase is not loop invariant.
636 if (L->contains(IP->getParent()))
637 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
638 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
639 InsertLoop = InsertLoop->getParentLoop();
642 // If there is no immediate value, skip the next part.
644 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
646 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
648 // If we are inserting the base and imm values in the same block, make sure to
649 // adjust the IP position if insertion reused a result.
650 if (IP == BaseInsertPt)
651 IP = Rewriter.getInsertionPoint();
653 // Always emit the immediate (if non-zero) into the same block as the user.
654 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
655 return Rewriter.expandCodeFor(NewValSCEV, IP);
660 // Once we rewrite the code to insert the new IVs we want, update the
661 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
662 // to it. NewBasePt is the last instruction which contributes to the
663 // value of NewBase in the case that it's a diffferent instruction from
664 // the PHI that NewBase is computed from, or null otherwise.
666 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
667 Instruction *NewBasePt,
668 SCEVExpander &Rewriter, Loop *L, Pass *P,
669 SmallVectorImpl<Instruction*> &DeadInsts){
670 if (!isa<PHINode>(Inst)) {
671 // By default, insert code at the user instruction.
672 BasicBlock::iterator InsertPt = Inst;
674 // However, if the Operand is itself an instruction, the (potentially
675 // complex) inserted code may be shared by many users. Because of this, we
676 // want to emit code for the computation of the operand right before its old
677 // computation. This is usually safe, because we obviously used to use the
678 // computation when it was computed in its current block. However, in some
679 // cases (e.g. use of a post-incremented induction variable) the NewBase
680 // value will be pinned to live somewhere after the original computation.
681 // In this case, we have to back off.
683 // If this is a use outside the loop (which means after, since it is based
684 // on a loop indvar) we use the post-incremented value, so that we don't
685 // artificially make the preinc value live out the bottom of the loop.
686 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
687 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
688 InsertPt = NewBasePt;
690 } else if (Instruction *OpInst
691 = dyn_cast<Instruction>(OperandValToReplace)) {
693 while (isa<PHINode>(InsertPt)) ++InsertPt;
696 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
697 // Adjust the type back to match the Inst. Note that we can't use InsertPt
698 // here because the SCEVExpander may have inserted the instructions after
699 // that point, in its efforts to avoid inserting redundant expressions.
700 if (isa<PointerType>(OperandValToReplace->getType())) {
701 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
703 OperandValToReplace->getType());
705 // Replace the use of the operand Value with the new Phi we just created.
706 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
707 DOUT << " CHANGED: IMM =" << *Imm;
708 DOUT << " \tNEWBASE =" << *NewBase;
709 DOUT << " \tInst = " << *Inst;
713 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
714 // expression into each operand block that uses it. Note that PHI nodes can
715 // have multiple entries for the same predecessor. We use a map to make sure
716 // that a PHI node only has a single Value* for each predecessor (which also
717 // prevents us from inserting duplicate code in some blocks).
718 DenseMap<BasicBlock*, Value*> InsertedCode;
719 PHINode *PN = cast<PHINode>(Inst);
720 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
721 if (PN->getIncomingValue(i) == OperandValToReplace) {
722 // If the original expression is outside the loop, put the replacement
723 // code in the same place as the original expression,
724 // which need not be an immediate predecessor of this PHI. This way we
725 // need only one copy of it even if it is referenced multiple times in
726 // the PHI. We don't do this when the original expression is inside the
727 // loop because multiple copies sometimes do useful sinking of code in that
729 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
730 if (L->contains(OldLoc->getParent())) {
731 // If this is a critical edge, split the edge so that we do not insert the
732 // code on all predecessor/successor paths. We do this unless this is the
733 // canonical backedge for this loop, as this can make some inserted code
734 // be in an illegal position.
735 BasicBlock *PHIPred = PN->getIncomingBlock(i);
736 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
737 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
739 // First step, split the critical edge.
740 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
742 // Next step: move the basic block. In particular, if the PHI node
743 // is outside of the loop, and PredTI is in the loop, we want to
744 // move the block to be immediately before the PHI block, not
745 // immediately after PredTI.
746 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
747 BasicBlock *NewBB = PN->getIncomingBlock(i);
748 NewBB->moveBefore(PN->getParent());
751 // Splitting the edge can reduce the number of PHI entries we have.
752 e = PN->getNumIncomingValues();
755 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
757 // Insert the code into the end of the predecessor block.
758 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
759 PN->getIncomingBlock(i)->getTerminator() :
760 OldLoc->getParent()->getTerminator();
761 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
763 // Adjust the type back to match the PHI. Note that we can't use
764 // InsertPt here because the SCEVExpander may have inserted its
765 // instructions after that point, in its efforts to avoid inserting
766 // redundant expressions.
767 if (isa<PointerType>(PN->getType())) {
768 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
774 // Replace the use of the operand Value with the new Phi we just created.
775 PN->setIncomingValue(i, Code);
780 // PHI node might have become a constant value after SplitCriticalEdge.
781 DeadInsts.push_back(Inst);
783 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
787 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
788 /// mode, and does not need to be put in a register first.
789 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
790 const TargetLowering *TLI, bool HasBaseReg) {
791 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
792 int64_t VC = SC->getValue()->getSExtValue();
794 TargetLowering::AddrMode AM;
796 AM.HasBaseReg = HasBaseReg;
797 return TLI->isLegalAddressingMode(AM, UseTy);
799 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
800 return (VC > -(1 << 16) && VC < (1 << 16)-1);
804 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
805 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
806 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
807 Constant *Op0 = CE->getOperand(0);
808 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
809 TargetLowering::AddrMode AM;
811 AM.HasBaseReg = HasBaseReg;
812 return TLI->isLegalAddressingMode(AM, UseTy);
818 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
819 /// loop varying to the Imm operand.
820 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
821 Loop *L, ScalarEvolution *SE) {
822 if (Val->isLoopInvariant(L)) return; // Nothing to do.
824 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
825 std::vector<SCEVHandle> NewOps;
826 NewOps.reserve(SAE->getNumOperands());
828 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
829 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
830 // If this is a loop-variant expression, it must stay in the immediate
831 // field of the expression.
832 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
834 NewOps.push_back(SAE->getOperand(i));
838 Val = SE->getIntegerSCEV(0, Val->getType());
840 Val = SE->getAddExpr(NewOps);
841 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
842 // Try to pull immediates out of the start value of nested addrec's.
843 SCEVHandle Start = SARE->getStart();
844 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
846 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
848 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
850 // Otherwise, all of Val is variant, move the whole thing over.
851 Imm = SE->getAddExpr(Imm, Val);
852 Val = SE->getIntegerSCEV(0, Val->getType());
857 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
858 /// that can fit into the immediate field of instructions in the target.
859 /// Accumulate these immediate values into the Imm value.
860 static void MoveImmediateValues(const TargetLowering *TLI,
862 SCEVHandle &Val, SCEVHandle &Imm,
863 bool isAddress, Loop *L,
864 ScalarEvolution *SE) {
865 const Type *UseTy = User->getType();
866 if (StoreInst *SI = dyn_cast<StoreInst>(User))
867 UseTy = SI->getOperand(0)->getType();
869 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
870 std::vector<SCEVHandle> NewOps;
871 NewOps.reserve(SAE->getNumOperands());
873 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
874 SCEVHandle NewOp = SAE->getOperand(i);
875 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
877 if (!NewOp->isLoopInvariant(L)) {
878 // If this is a loop-variant expression, it must stay in the immediate
879 // field of the expression.
880 Imm = SE->getAddExpr(Imm, NewOp);
882 NewOps.push_back(NewOp);
887 Val = SE->getIntegerSCEV(0, Val->getType());
889 Val = SE->getAddExpr(NewOps);
891 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
892 // Try to pull immediates out of the start value of nested addrec's.
893 SCEVHandle Start = SARE->getStart();
894 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
896 if (Start != SARE->getStart()) {
897 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
899 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
902 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
903 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
904 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
905 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
907 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
908 SCEVHandle NewOp = SME->getOperand(1);
909 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
911 // If we extracted something out of the subexpressions, see if we can
913 if (NewOp != SME->getOperand(1)) {
914 // Scale SubImm up by "8". If the result is a target constant, we are
916 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
917 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
918 // Accumulate the immediate.
919 Imm = SE->getAddExpr(Imm, SubImm);
921 // Update what is left of 'Val'.
922 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
929 // Loop-variant expressions must stay in the immediate field of the
931 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
932 !Val->isLoopInvariant(L)) {
933 Imm = SE->getAddExpr(Imm, Val);
934 Val = SE->getIntegerSCEV(0, Val->getType());
938 // Otherwise, no immediates to move.
942 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
943 /// added together. This is used to reassociate common addition subexprs
944 /// together for maximal sharing when rewriting bases.
945 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
947 ScalarEvolution *SE) {
948 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
949 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
950 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
951 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
952 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
953 if (SARE->getOperand(0) == Zero) {
954 SubExprs.push_back(Expr);
956 // Compute the addrec with zero as its base.
957 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
958 Ops[0] = Zero; // Start with zero base.
959 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
962 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
964 } else if (!Expr->isZero()) {
966 SubExprs.push_back(Expr);
970 // This is logically local to the following function, but C++ says we have
971 // to make it file scope.
972 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
974 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
975 /// the Uses, removing any common subexpressions, except that if all such
976 /// subexpressions can be folded into an addressing mode for all uses inside
977 /// the loop (this case is referred to as "free" in comments herein) we do
978 /// not remove anything. This looks for things like (a+b+c) and
979 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
980 /// is *removed* from the Bases and returned.
982 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
983 ScalarEvolution *SE, Loop *L,
984 const TargetLowering *TLI) {
985 unsigned NumUses = Uses.size();
987 // Only one use? This is a very common case, so we handle it specially and
989 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
990 SCEVHandle Result = Zero;
991 SCEVHandle FreeResult = Zero;
993 // If the use is inside the loop, use its base, regardless of what it is:
994 // it is clearly shared across all the IV's. If the use is outside the loop
995 // (which means after it) we don't want to factor anything *into* the loop,
996 // so just use 0 as the base.
997 if (L->contains(Uses[0].Inst->getParent()))
998 std::swap(Result, Uses[0].Base);
1002 // To find common subexpressions, count how many of Uses use each expression.
1003 // If any subexpressions are used Uses.size() times, they are common.
1004 // Also track whether all uses of each expression can be moved into an
1005 // an addressing mode "for free"; such expressions are left within the loop.
1006 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1007 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
1009 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
1010 // order we see them.
1011 std::vector<SCEVHandle> UniqueSubExprs;
1013 std::vector<SCEVHandle> SubExprs;
1014 unsigned NumUsesInsideLoop = 0;
1015 for (unsigned i = 0; i != NumUses; ++i) {
1016 // If the user is outside the loop, just ignore it for base computation.
1017 // Since the user is outside the loop, it must be *after* the loop (if it
1018 // were before, it could not be based on the loop IV). We don't want users
1019 // after the loop to affect base computation of values *inside* the loop,
1020 // because we can always add their offsets to the result IV after the loop
1021 // is done, ensuring we get good code inside the loop.
1022 if (!L->contains(Uses[i].Inst->getParent()))
1024 NumUsesInsideLoop++;
1026 // If the base is zero (which is common), return zero now, there are no
1027 // CSEs we can find.
1028 if (Uses[i].Base == Zero) return Zero;
1030 // If this use is as an address we may be able to put CSEs in the addressing
1031 // mode rather than hoisting them.
1032 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1033 // We may need the UseTy below, but only when isAddrUse, so compute it
1034 // only in that case.
1035 const Type *UseTy = 0;
1037 UseTy = Uses[i].Inst->getType();
1038 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1039 UseTy = SI->getOperand(0)->getType();
1042 // Split the expression into subexprs.
1043 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1044 // Add one to SubExpressionUseData.Count for each subexpr present, and
1045 // if the subexpr is not a valid immediate within an addressing mode use,
1046 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1047 // hoist these out of the loop (if they are common to all uses).
1048 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1049 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1050 UniqueSubExprs.push_back(SubExprs[j]);
1051 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1052 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1057 // Now that we know how many times each is used, build Result. Iterate over
1058 // UniqueSubexprs so that we have a stable ordering.
1059 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1060 std::map<SCEVHandle, SubExprUseData>::iterator I =
1061 SubExpressionUseData.find(UniqueSubExprs[i]);
1062 assert(I != SubExpressionUseData.end() && "Entry not found?");
1063 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1064 if (I->second.notAllUsesAreFree)
1065 Result = SE->getAddExpr(Result, I->first);
1067 FreeResult = SE->getAddExpr(FreeResult, I->first);
1069 // Remove non-cse's from SubExpressionUseData.
1070 SubExpressionUseData.erase(I);
1073 if (FreeResult != Zero) {
1074 // We have some subexpressions that can be subsumed into addressing
1075 // modes in every use inside the loop. However, it's possible that
1076 // there are so many of them that the combined FreeResult cannot
1077 // be subsumed, or that the target cannot handle both a FreeResult
1078 // and a Result in the same instruction (for example because it would
1079 // require too many registers). Check this.
1080 for (unsigned i=0; i<NumUses; ++i) {
1081 if (!L->contains(Uses[i].Inst->getParent()))
1083 // We know this is an addressing mode use; if there are any uses that
1084 // are not, FreeResult would be Zero.
1085 const Type *UseTy = Uses[i].Inst->getType();
1086 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1087 UseTy = SI->getOperand(0)->getType();
1088 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1089 // FIXME: could split up FreeResult into pieces here, some hoisted
1090 // and some not. There is no obvious advantage to this.
1091 Result = SE->getAddExpr(Result, FreeResult);
1098 // If we found no CSE's, return now.
1099 if (Result == Zero) return Result;
1101 // If we still have a FreeResult, remove its subexpressions from
1102 // SubExpressionUseData. This means they will remain in the use Bases.
1103 if (FreeResult != Zero) {
1104 SeparateSubExprs(SubExprs, FreeResult, SE);
1105 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1106 std::map<SCEVHandle, SubExprUseData>::iterator I =
1107 SubExpressionUseData.find(SubExprs[j]);
1108 SubExpressionUseData.erase(I);
1113 // Otherwise, remove all of the CSE's we found from each of the base values.
1114 for (unsigned i = 0; i != NumUses; ++i) {
1115 // Uses outside the loop don't necessarily include the common base, but
1116 // the final IV value coming into those uses does. Instead of trying to
1117 // remove the pieces of the common base, which might not be there,
1118 // subtract off the base to compensate for this.
1119 if (!L->contains(Uses[i].Inst->getParent())) {
1120 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1124 // Split the expression into subexprs.
1125 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1127 // Remove any common subexpressions.
1128 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1129 if (SubExpressionUseData.count(SubExprs[j])) {
1130 SubExprs.erase(SubExprs.begin()+j);
1134 // Finally, add the non-shared expressions together.
1135 if (SubExprs.empty())
1136 Uses[i].Base = Zero;
1138 Uses[i].Base = SE->getAddExpr(SubExprs);
1145 /// ValidStride - Check whether the given Scale is valid for all loads and
1146 /// stores in UsersToProcess.
1148 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1150 const std::vector<BasedUser>& UsersToProcess) {
1154 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1155 // If this is a load or other access, pass the type of the access in.
1156 const Type *AccessTy = Type::VoidTy;
1157 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1158 AccessTy = SI->getOperand(0)->getType();
1159 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1160 AccessTy = LI->getType();
1161 else if (isa<PHINode>(UsersToProcess[i].Inst))
1164 TargetLowering::AddrMode AM;
1165 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1166 AM.BaseOffs = SC->getValue()->getSExtValue();
1167 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1170 // If load[imm+r*scale] is illegal, bail out.
1171 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1177 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1179 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1183 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1185 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1186 !(isa<PointerType>(Ty2) &&
1187 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1188 !(isa<PointerType>(Ty1) &&
1189 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1192 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1193 /// of a previous stride and it is a legal value for the target addressing
1194 /// mode scale component and optional base reg. This allows the users of
1195 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1196 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1198 /// If all uses are outside the loop, we don't require that all multiplies
1199 /// be folded into the addressing mode; a multiply (executed once) outside
1200 /// the loop is better than another IV within. Well, usually.
1201 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1202 bool AllUsesAreAddresses,
1203 bool AllUsesAreOutsideLoop,
1204 const SCEVHandle &Stride,
1205 IVExpr &IV, const Type *Ty,
1206 const std::vector<BasedUser>& UsersToProcess) {
1207 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1208 int64_t SInt = SC->getValue()->getSExtValue();
1209 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1211 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1212 IVsByStride.find(StrideOrder[NewStride]);
1213 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1215 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1216 if (SI->first != Stride &&
1217 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1219 int64_t Scale = SInt / SSInt;
1220 // Check that this stride is valid for all the types used for loads and
1221 // stores; if it can be used for some and not others, we might as well use
1222 // the original stride everywhere, since we have to create the IV for it
1223 // anyway. If the scale is 1, then we don't need to worry about folding
1226 (AllUsesAreAddresses &&
1227 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1228 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1229 IE = SI->second.IVs.end(); II != IE; ++II)
1230 // FIXME: Only handle base == 0 for now.
1231 // Only reuse previous IV if it would not require a type conversion.
1232 if (II->Base->isZero() &&
1233 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1235 return SE->getIntegerSCEV(Scale, Stride->getType());
1238 } else if (AllUsesAreOutsideLoop) {
1239 // Accept nonconstant strides here; it is really really right to substitute
1240 // an existing IV if we can.
1241 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1243 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1244 IVsByStride.find(StrideOrder[NewStride]);
1245 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1247 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1248 if (SI->first != Stride && SSInt != 1)
1250 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1251 IE = SI->second.IVs.end(); II != IE; ++II)
1252 // Accept nonzero base here.
1253 // Only reuse previous IV if it would not require a type conversion.
1254 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1259 // Special case, old IV is -1*x and this one is x. Can treat this one as
1261 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1263 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1264 IVsByStride.find(StrideOrder[NewStride]);
1265 if (SI == IVsByStride.end())
1267 if (SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1268 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1269 if (Stride == ME->getOperand(1) &&
1270 SC->getValue()->getSExtValue() == -1LL)
1271 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1272 IE = SI->second.IVs.end(); II != IE; ++II)
1273 // Accept nonzero base here.
1274 // Only reuse previous IV if it would not require type conversion.
1275 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1277 return SE->getIntegerSCEV(-1LL, Stride->getType());
1281 return SE->getIntegerSCEV(0, Stride->getType());
1284 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1285 /// returns true if Val's isUseOfPostIncrementedValue is true.
1286 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1287 return Val.isUseOfPostIncrementedValue;
1290 /// isNonConstantNegative - Return true if the specified scev is negated, but
1292 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1293 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1294 if (!Mul) return false;
1296 // If there is a constant factor, it will be first.
1297 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1298 if (!SC) return false;
1300 // Return true if the value is negative, this matches things like (-42 * V).
1301 return SC->getValue()->getValue().isNegative();
1304 // CollectIVUsers - Transform our list of users and offsets to a bit more
1305 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1306 // of the strided accesses, as well as the old information from Uses. We
1307 // progressively move information from the Base field to the Imm field, until
1308 // we eventually have the full access expression to rewrite the use.
1309 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1310 IVUsersOfOneStride &Uses,
1312 bool &AllUsesAreAddresses,
1313 bool &AllUsesAreOutsideLoop,
1314 std::vector<BasedUser> &UsersToProcess) {
1315 UsersToProcess.reserve(Uses.Users.size());
1316 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1317 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1319 // Move any loop variant operands from the offset field to the immediate
1320 // field of the use, so that we don't try to use something before it is
1322 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1323 UsersToProcess.back().Imm, L, SE);
1324 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1325 "Base value is not loop invariant!");
1328 // We now have a whole bunch of uses of like-strided induction variables, but
1329 // they might all have different bases. We want to emit one PHI node for this
1330 // stride which we fold as many common expressions (between the IVs) into as
1331 // possible. Start by identifying the common expressions in the base values
1332 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1333 // "A+B"), emit it to the preheader, then remove the expression from the
1334 // UsersToProcess base values.
1335 SCEVHandle CommonExprs =
1336 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1338 // Next, figure out what we can represent in the immediate fields of
1339 // instructions. If we can represent anything there, move it to the imm
1340 // fields of the BasedUsers. We do this so that it increases the commonality
1341 // of the remaining uses.
1342 unsigned NumPHI = 0;
1343 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1344 // If the user is not in the current loop, this means it is using the exit
1345 // value of the IV. Do not put anything in the base, make sure it's all in
1346 // the immediate field to allow as much factoring as possible.
1347 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1348 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1349 UsersToProcess[i].Base);
1350 UsersToProcess[i].Base =
1351 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1354 // Addressing modes can be folded into loads and stores. Be careful that
1355 // the store is through the expression, not of the expression though.
1357 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1358 UsersToProcess[i].OperandValToReplace);
1359 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1364 // Not all uses are outside the loop.
1365 AllUsesAreOutsideLoop = false;
1367 // If this use isn't an address, then not all uses are addresses.
1368 if (!isAddress && !isPHI)
1369 AllUsesAreAddresses = false;
1371 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1372 UsersToProcess[i].Imm, isAddress, L, SE);
1376 // If one of the use if a PHI node and all other uses are addresses, still
1377 // allow iv reuse. Essentially we are trading one constant multiplication
1378 // for one fewer iv.
1380 AllUsesAreAddresses = false;
1385 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1386 /// stride of IV. All of the users may have different starting values, and this
1387 /// may not be the only stride (we know it is if isOnlyStride is true).
1388 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1389 IVUsersOfOneStride &Uses,
1391 bool isOnlyStride) {
1392 // If all the users are moved to another stride, then there is nothing to do.
1393 if (Uses.Users.empty())
1396 // Keep track if every use in UsersToProcess is an address. If they all are,
1397 // we may be able to rewrite the entire collection of them in terms of a
1398 // smaller-stride IV.
1399 bool AllUsesAreAddresses = true;
1401 // Keep track if every use of a single stride is outside the loop. If so,
1402 // we want to be more aggressive about reusing a smaller-stride IV; a
1403 // multiply outside the loop is better than another IV inside. Well, usually.
1404 bool AllUsesAreOutsideLoop = true;
1406 // Transform our list of users and offsets to a bit more complex table. In
1407 // this new vector, each 'BasedUser' contains 'Base' the base of the
1408 // strided accessas well as the old information from Uses. We progressively
1409 // move information from the Base field to the Imm field, until we eventually
1410 // have the full access expression to rewrite the use.
1411 std::vector<BasedUser> UsersToProcess;
1412 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1413 AllUsesAreOutsideLoop,
1416 // If we managed to find some expressions in common, we'll need to carry
1417 // their value in a register and add it in for each use. This will take up
1418 // a register operand, which potentially restricts what stride values are
1420 bool HaveCommonExprs = !CommonExprs->isZero();
1422 // If all uses are addresses, check if it is possible to reuse an IV with a
1423 // stride that is a factor of this stride. And that the multiple is a number
1424 // that can be encoded in the scale field of the target addressing mode. And
1425 // that we will have a valid instruction after this substition, including the
1426 // immediate field, if any.
1427 PHINode *NewPHI = NULL;
1429 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1430 SE->getIntegerSCEV(0, Type::Int32Ty),
1432 SCEVHandle RewriteFactor =
1433 CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1434 AllUsesAreOutsideLoop,
1435 Stride, ReuseIV, CommonExprs->getType(),
1437 if (!isa<SCEVConstant>(RewriteFactor) ||
1438 !cast<SCEVConstant>(RewriteFactor)->isZero()) {
1439 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1440 << " and BASE " << *ReuseIV.Base << " :\n";
1441 NewPHI = ReuseIV.PHI;
1442 IncV = ReuseIV.IncV;
1445 const Type *ReplacedTy = CommonExprs->getType();
1447 // Now that we know what we need to do, insert the PHI node itself.
1449 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1450 << *Stride << " and BASE " << *CommonExprs << ": ";
1452 SCEVExpander Rewriter(*SE, *LI);
1453 SCEVExpander PreheaderRewriter(*SE, *LI);
1455 BasicBlock *Preheader = L->getLoopPreheader();
1456 Instruction *PreInsertPt = Preheader->getTerminator();
1457 Instruction *PhiInsertBefore = L->getHeader()->begin();
1459 BasicBlock *LatchBlock = L->getLoopLatch();
1462 // Emit the initial base value into the loop preheader.
1464 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1466 if (isa<SCEVConstant>(RewriteFactor) &&
1467 cast<SCEVConstant>(RewriteFactor)->isZero()) {
1468 // Create a new Phi for this base, and stick it in the loop header.
1469 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1472 // Add common base to the new Phi node.
1473 NewPHI->addIncoming(CommonBaseV, Preheader);
1475 // If the stride is negative, insert a sub instead of an add for the
1477 bool isNegative = isNonConstantNegative(Stride);
1478 SCEVHandle IncAmount = Stride;
1480 IncAmount = SE->getNegativeSCEV(Stride);
1482 // Insert the stride into the preheader.
1483 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1484 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1486 // Emit the increment of the base value before the terminator of the loop
1487 // latch block, and add it to the Phi node.
1488 SCEVHandle IncExp = SE->getUnknown(StrideV);
1490 IncExp = SE->getNegativeSCEV(IncExp);
1491 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1493 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1494 IncV->setName(NewPHI->getName()+".inc");
1495 NewPHI->addIncoming(IncV, LatchBlock);
1497 // Remember this in case a later stride is multiple of this.
1498 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1500 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1502 Constant *C = dyn_cast<Constant>(CommonBaseV);
1504 (!C->isNullValue() &&
1505 !fitsInAddressMode(SE->getUnknown(CommonBaseV), ReplacedTy,
1507 // We want the common base emitted into the preheader! This is just
1508 // using cast as a copy so BitCast (no-op cast) is appropriate
1509 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1510 "commonbase", PreInsertPt);
1514 // We want to emit code for users inside the loop first. To do this, we
1515 // rearrange BasedUser so that the entries at the end have
1516 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1517 // vector (so we handle them first).
1518 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1519 PartitionByIsUseOfPostIncrementedValue);
1521 // Sort this by base, so that things with the same base are handled
1522 // together. By partitioning first and stable-sorting later, we are
1523 // guaranteed that within each base we will pop off users from within the
1524 // loop before users outside of the loop with a particular base.
1526 // We would like to use stable_sort here, but we can't. The problem is that
1527 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1528 // we don't have anything to do a '<' comparison on. Because we think the
1529 // number of uses is small, do a horrible bubble sort which just relies on
1531 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1532 // Get a base value.
1533 SCEVHandle Base = UsersToProcess[i].Base;
1535 // Compact everything with this base to be consecutive with this one.
1536 for (unsigned j = i+1; j != e; ++j) {
1537 if (UsersToProcess[j].Base == Base) {
1538 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1544 // Process all the users now. This outer loop handles all bases, the inner
1545 // loop handles all users of a particular base.
1546 while (!UsersToProcess.empty()) {
1547 SCEVHandle Base = UsersToProcess.back().Base;
1549 // Emit the code for Base into the preheader.
1550 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1552 DOUT << " INSERTING code for BASE = " << *Base << ":";
1553 if (BaseV->hasName())
1554 DOUT << " Result value name = %" << BaseV->getNameStr();
1557 // If BaseV is a constant other than 0, make sure that it gets inserted into
1558 // the preheader, instead of being forward substituted into the uses. We do
1559 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1561 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1562 if (!C->isNullValue() && !fitsInAddressMode(Base, ReplacedTy,
1564 // We want this constant emitted into the preheader! This is just
1565 // using cast as a copy so BitCast (no-op cast) is appropriate
1566 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1571 // Emit the code to add the immediate offset to the Phi value, just before
1572 // the instructions that we identified as using this stride and base.
1574 // FIXME: Use emitted users to emit other users.
1575 BasedUser &User = UsersToProcess.back();
1577 // If this instruction wants to use the post-incremented value, move it
1578 // after the post-inc and use its value instead of the PHI.
1579 Value *RewriteOp = NewPHI;
1580 if (User.isUseOfPostIncrementedValue) {
1583 // If this user is in the loop, make sure it is the last thing in the
1584 // loop to ensure it is dominated by the increment.
1585 if (L->contains(User.Inst->getParent()))
1586 User.Inst->moveBefore(LatchBlock->getTerminator());
1588 if (RewriteOp->getType() != ReplacedTy) {
1589 Instruction::CastOps opcode = Instruction::Trunc;
1590 if (ReplacedTy->getPrimitiveSizeInBits() ==
1591 RewriteOp->getType()->getPrimitiveSizeInBits())
1592 opcode = Instruction::BitCast;
1593 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1596 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1598 // If we had to insert new instructions for RewriteOp, we have to
1599 // consider that they may not have been able to end up immediately
1600 // next to RewriteOp, because non-PHI instructions may never precede
1601 // PHI instructions in a block. In this case, remember where the last
1602 // instruction was inserted so that if we're replacing a different
1603 // PHI node, we can use the later point to expand the final
1605 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1606 if (RewriteOp == NewPHI) NewBasePt = 0;
1608 // Clear the SCEVExpander's expression map so that we are guaranteed
1609 // to have the code emitted where we expect it.
1612 // If we are reusing the iv, then it must be multiplied by a constant
1613 // factor take advantage of addressing mode scale component.
1614 if (!isa<SCEVConstant>(RewriteFactor) ||
1615 !cast<SCEVConstant>(RewriteFactor)->isZero()) {
1616 // If we're reusing an IV with a nonzero base (currently this happens
1617 // only when all reuses are outside the loop) subtract that base here.
1618 // The base has been used to initialize the PHI node but we don't want
1620 if (!ReuseIV.Base->isZero())
1621 RewriteExpr = SE->getMinusSCEV(RewriteExpr, ReuseIV.Base);
1623 // Multiply old variable, with base removed, by new scale factor.
1624 RewriteExpr = SE->getMulExpr(RewriteFactor,
1627 // The common base is emitted in the loop preheader. But since we
1628 // are reusing an IV, it has not been used to initialize the PHI node.
1629 // Add it to the expression used to rewrite the uses.
1630 if (!isa<ConstantInt>(CommonBaseV) ||
1631 !cast<ConstantInt>(CommonBaseV)->isZero())
1632 RewriteExpr = SE->getAddExpr(RewriteExpr,
1633 SE->getUnknown(CommonBaseV));
1636 // Now that we know what we need to do, insert code before User for the
1637 // immediate and any loop-variant expressions.
1638 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1639 // Add BaseV to the PHI value if needed.
1640 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1642 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1646 // Mark old value we replaced as possibly dead, so that it is eliminated
1647 // if we just replaced the last use of that value.
1648 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1650 UsersToProcess.pop_back();
1653 // If there are any more users to process with the same base, process them
1654 // now. We sorted by base above, so we just have to check the last elt.
1655 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1656 // TODO: Next, find out which base index is the most common, pull it out.
1659 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1660 // different starting values, into different PHIs.
1663 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1664 /// set the IV user and stride information and return true, otherwise return
1666 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1667 const SCEVHandle *&CondStride) {
1668 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1670 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1671 IVUsesByStride.find(StrideOrder[Stride]);
1672 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1674 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1675 E = SI->second.Users.end(); UI != E; ++UI)
1676 if (UI->User == Cond) {
1677 // NOTE: we could handle setcc instructions with multiple uses here, but
1678 // InstCombine does it as well for simple uses, it's not clear that it
1679 // occurs enough in real life to handle.
1681 CondStride = &SI->first;
1689 // Constant strides come first which in turns are sorted by their absolute
1690 // values. If absolute values are the same, then positive strides comes first.
1692 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1693 struct StrideCompare {
1694 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1695 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1696 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1698 int64_t LV = LHSC->getValue()->getSExtValue();
1699 int64_t RV = RHSC->getValue()->getSExtValue();
1700 uint64_t ALV = (LV < 0) ? -LV : LV;
1701 uint64_t ARV = (RV < 0) ? -RV : RV;
1707 return (LHSC && !RHSC);
1712 /// ChangeCompareStride - If a loop termination compare instruction is the
1713 /// only use of its stride, and the compaison is against a constant value,
1714 /// try eliminate the stride by moving the compare instruction to another
1715 /// stride and change its constant operand accordingly. e.g.
1721 /// if (v2 < 10) goto loop
1726 /// if (v1 < 30) goto loop
1727 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1728 IVStrideUse* &CondUse,
1729 const SCEVHandle* &CondStride) {
1730 if (StrideOrder.size() < 2 ||
1731 IVUsesByStride[*CondStride].Users.size() != 1)
1733 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1734 if (!SC) return Cond;
1735 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1736 if (!C) return Cond;
1738 ICmpInst::Predicate Predicate = Cond->getPredicate();
1739 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1740 int64_t CmpVal = C->getValue().getSExtValue();
1741 unsigned BitWidth = C->getValue().getBitWidth();
1742 uint64_t SignBit = 1ULL << (BitWidth-1);
1743 const Type *CmpTy = C->getType();
1744 const Type *NewCmpTy = NULL;
1745 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1746 unsigned NewTyBits = 0;
1747 int64_t NewCmpVal = CmpVal;
1748 SCEVHandle *NewStride = NULL;
1749 Value *NewIncV = NULL;
1752 // Check stride constant and the comparision constant signs to detect
1754 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1757 // Look for a suitable stride / iv as replacement.
1758 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1759 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1760 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1761 IVUsesByStride.find(StrideOrder[i]);
1762 if (!isa<SCEVConstant>(SI->first))
1764 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1765 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1768 Scale = SSInt / CmpSSInt;
1769 NewCmpVal = CmpVal * Scale;
1770 APInt Mul = APInt(BitWidth, NewCmpVal);
1771 // Check for overflow.
1772 if (Mul.getSExtValue() != NewCmpVal) {
1777 // Watch out for overflow.
1778 if (ICmpInst::isSignedPredicate(Predicate) &&
1779 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1782 if (NewCmpVal != CmpVal) {
1783 // Pick the best iv to use trying to avoid a cast.
1785 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1786 E = SI->second.Users.end(); UI != E; ++UI) {
1787 NewIncV = UI->OperandValToReplace;
1788 if (NewIncV->getType() == CmpTy)
1796 NewCmpTy = NewIncV->getType();
1797 NewTyBits = isa<PointerType>(NewCmpTy)
1798 ? UIntPtrTy->getPrimitiveSizeInBits()
1799 : NewCmpTy->getPrimitiveSizeInBits();
1800 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1801 // Check if it is possible to rewrite it using
1802 // an iv / stride of a smaller integer type.
1803 bool TruncOk = false;
1804 if (NewCmpTy->isInteger()) {
1805 unsigned Bits = NewTyBits;
1806 if (ICmpInst::isSignedPredicate(Predicate))
1808 uint64_t Mask = (1ULL << Bits) - 1;
1809 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1818 // Don't rewrite if use offset is non-constant and the new type is
1819 // of a different type.
1820 // FIXME: too conservative?
1821 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1826 bool AllUsesAreAddresses = true;
1827 bool AllUsesAreOutsideLoop = true;
1828 std::vector<BasedUser> UsersToProcess;
1829 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1830 AllUsesAreAddresses,
1831 AllUsesAreOutsideLoop,
1833 // Avoid rewriting the compare instruction with an iv of new stride
1834 // if it's likely the new stride uses will be rewritten using the
1835 if (AllUsesAreAddresses &&
1836 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1841 // If scale is negative, use swapped predicate unless it's testing
1843 if (Scale < 0 && !Cond->isEquality())
1844 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1846 NewStride = &StrideOrder[i];
1851 // Forgo this transformation if it the increment happens to be
1852 // unfortunately positioned after the condition, and the condition
1853 // has multiple uses which prevent it from being moved immediately
1854 // before the branch. See
1855 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1856 // for an example of this situation.
1857 if (!Cond->hasOneUse()) {
1858 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1864 if (NewCmpVal != CmpVal) {
1865 // Create a new compare instruction using new stride / iv.
1866 ICmpInst *OldCond = Cond;
1868 if (!isa<PointerType>(NewCmpTy))
1869 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1871 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1872 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1874 // Insert new compare instruction.
1875 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1876 L->getHeader()->getName() + ".termcond",
1879 // Remove the old compare instruction. The old indvar is probably dead too.
1880 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
1881 SE->deleteValueFromRecords(OldCond);
1882 OldCond->replaceAllUsesWith(Cond);
1883 OldCond->eraseFromParent();
1885 IVUsesByStride[*CondStride].Users.pop_back();
1886 SCEVHandle NewOffset = TyBits == NewTyBits
1887 ? SE->getMulExpr(CondUse->Offset,
1888 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1889 : SE->getConstant(ConstantInt::get(NewCmpTy,
1890 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1891 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1892 CondUse = &IVUsesByStride[*NewStride].Users.back();
1893 CondStride = NewStride;
1900 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1901 /// an smax computation.
1903 /// This is a narrow solution to a specific, but acute, problem. For loops
1909 /// } while (++i < n);
1911 /// where the comparison is signed, the trip count isn't just 'n', because
1912 /// 'n' could be negative. And unfortunately this can come up even for loops
1913 /// where the user didn't use a C do-while loop. For example, seemingly
1914 /// well-behaved top-test loops will commonly be lowered like this:
1920 /// } while (++i < n);
1923 /// and then it's possible for subsequent optimization to obscure the if
1924 /// test in such a way that indvars can't find it.
1926 /// When indvars can't find the if test in loops like this, it creates a
1927 /// signed-max expression, which allows it to give the loop a canonical
1928 /// induction variable:
1931 /// smax = n < 1 ? 1 : n;
1934 /// } while (++i != smax);
1936 /// Canonical induction variables are necessary because the loop passes
1937 /// are designed around them. The most obvious example of this is the
1938 /// LoopInfo analysis, which doesn't remember trip count values. It
1939 /// expects to be able to rediscover the trip count each time it is
1940 /// needed, and it does this using a simple analyis that only succeeds if
1941 /// the loop has a canonical induction variable.
1943 /// However, when it comes time to generate code, the maximum operation
1944 /// can be quite costly, especially if it's inside of an outer loop.
1946 /// This function solves this problem by detecting this type of loop and
1947 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1948 /// the instructions for the maximum computation.
1950 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1951 IVStrideUse* &CondUse) {
1952 // Check that the loop matches the pattern we're looking for.
1953 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1954 Cond->getPredicate() != CmpInst::ICMP_NE)
1957 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1958 if (!Sel || !Sel->hasOneUse()) return Cond;
1960 SCEVHandle IterationCount = SE->getIterationCount(L);
1961 if (isa<SCEVCouldNotCompute>(IterationCount))
1963 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1965 // Adjust for an annoying getIterationCount quirk.
1966 IterationCount = SE->getAddExpr(IterationCount, One);
1968 // Check for a max calculation that matches the pattern.
1969 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1970 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1972 SCEVHandle SMaxLHS = SMax->getOperand(0);
1973 SCEVHandle SMaxRHS = SMax->getOperand(1);
1974 if (!SMaxLHS || SMaxLHS != One) return Cond;
1976 // Check the relevant induction variable for conformance to
1978 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1979 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1980 if (!AR || !AR->isAffine() ||
1981 AR->getStart() != One ||
1982 AR->getStepRecurrence(*SE) != One)
1985 // Check the right operand of the select, and remember it, as it will
1986 // be used in the new comparison instruction.
1988 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1989 NewRHS = Sel->getOperand(1);
1990 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1991 NewRHS = Sel->getOperand(2);
1992 if (!NewRHS) return Cond;
1994 // Ok, everything looks ok to change the condition into an SLT or SGE and
1995 // delete the max calculation.
1997 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2000 Cond->getOperand(0), NewRHS, "scmp", Cond);
2002 // Delete the max calculation instructions.
2003 SE->deleteValueFromRecords(Cond);
2004 Cond->replaceAllUsesWith(NewCond);
2005 Cond->eraseFromParent();
2006 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2007 SE->deleteValueFromRecords(Sel);
2008 Sel->eraseFromParent();
2009 if (Cmp->use_empty()) {
2010 SE->deleteValueFromRecords(Cmp);
2011 Cmp->eraseFromParent();
2013 CondUse->User = NewCond;
2017 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2018 /// inside the loop then try to eliminate the cast opeation.
2019 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2021 SCEVHandle IterationCount = SE->getIterationCount(L);
2022 if (isa<SCEVCouldNotCompute>(IterationCount))
2025 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2027 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2028 IVUsesByStride.find(StrideOrder[Stride]);
2029 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2030 if (!isa<SCEVConstant>(SI->first))
2033 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2034 E = SI->second.Users.end(); UI != E; /* empty */) {
2035 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2037 Instruction *ShadowUse = CandidateUI->User;
2038 const Type *DestTy = NULL;
2040 /* If shadow use is a int->float cast then insert a second IV
2041 to eliminate this cast.
2043 for (unsigned i = 0; i < n; ++i)
2049 for (unsigned i = 0; i < n; ++i, ++d)
2052 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2053 DestTy = UCast->getDestTy();
2054 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2055 DestTy = SCast->getDestTy();
2056 if (!DestTy) continue;
2059 /* If target does not support DestTy natively then do not apply
2060 this transformation. */
2061 MVT DVT = TLI->getValueType(DestTy);
2062 if (!TLI->isTypeLegal(DVT)) continue;
2065 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2067 if (PH->getNumIncomingValues() != 2) continue;
2069 const Type *SrcTy = PH->getType();
2070 int Mantissa = DestTy->getFPMantissaWidth();
2071 if (Mantissa == -1) continue;
2072 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
2075 unsigned Entry, Latch;
2076 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2084 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2085 if (!Init) continue;
2086 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2088 BinaryOperator *Incr =
2089 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2090 if (!Incr) continue;
2091 if (Incr->getOpcode() != Instruction::Add
2092 && Incr->getOpcode() != Instruction::Sub)
2095 /* Initialize new IV, double d = 0.0 in above example. */
2096 ConstantInt *C = NULL;
2097 if (Incr->getOperand(0) == PH)
2098 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2099 else if (Incr->getOperand(1) == PH)
2100 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2106 /* Add new PHINode. */
2107 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2109 /* create new increment. '++d' in above example. */
2110 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2111 BinaryOperator *NewIncr =
2112 BinaryOperator::Create(Incr->getOpcode(),
2113 NewPH, CFP, "IV.S.next.", Incr);
2115 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2116 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2118 /* Remove cast operation */
2119 SE->deleteValueFromRecords(ShadowUse);
2120 ShadowUse->replaceAllUsesWith(NewPH);
2121 ShadowUse->eraseFromParent();
2122 SI->second.Users.erase(CandidateUI);
2129 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2130 // uses in the loop, look to see if we can eliminate some, in favor of using
2131 // common indvars for the different uses.
2132 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2133 // TODO: implement optzns here.
2135 OptimizeShadowIV(L);
2137 // Finally, get the terminating condition for the loop if possible. If we
2138 // can, we want to change it to use a post-incremented version of its
2139 // induction variable, to allow coalescing the live ranges for the IV into
2140 // one register value.
2141 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2142 BasicBlock *Preheader = L->getLoopPreheader();
2143 BasicBlock *LatchBlock =
2144 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2145 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2146 if (!TermBr || TermBr->isUnconditional() ||
2147 !isa<ICmpInst>(TermBr->getCondition()))
2149 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2151 // Search IVUsesByStride to find Cond's IVUse if there is one.
2152 IVStrideUse *CondUse = 0;
2153 const SCEVHandle *CondStride = 0;
2155 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2156 return; // setcc doesn't use the IV.
2158 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2159 // being unable to find a sufficient guard, for example), change the loop
2160 // comparison to use SLT instead of NE.
2161 Cond = OptimizeSMax(L, Cond, CondUse);
2163 // If possible, change stride and operands of the compare instruction to
2164 // eliminate one stride.
2165 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2167 // It's possible for the setcc instruction to be anywhere in the loop, and
2168 // possible for it to have multiple users. If it is not immediately before
2169 // the latch block branch, move it.
2170 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2171 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2172 Cond->moveBefore(TermBr);
2174 // Otherwise, clone the terminating condition and insert into the loopend.
2175 Cond = cast<ICmpInst>(Cond->clone());
2176 Cond->setName(L->getHeader()->getName() + ".termcond");
2177 LatchBlock->getInstList().insert(TermBr, Cond);
2179 // Clone the IVUse, as the old use still exists!
2180 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2181 CondUse->OperandValToReplace);
2182 CondUse = &IVUsesByStride[*CondStride].Users.back();
2186 // If we get to here, we know that we can transform the setcc instruction to
2187 // use the post-incremented version of the IV, allowing us to coalesce the
2188 // live ranges for the IV correctly.
2189 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2190 CondUse->isUseOfPostIncrementedValue = true;
2194 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2196 LI = &getAnalysis<LoopInfo>();
2197 DT = &getAnalysis<DominatorTree>();
2198 SE = &getAnalysis<ScalarEvolution>();
2199 TD = &getAnalysis<TargetData>();
2200 UIntPtrTy = TD->getIntPtrType();
2203 // Find all uses of induction variables in this loop, and categorize
2204 // them by stride. Start by finding all of the PHI nodes in the header for
2205 // this loop. If they are induction variables, inspect their uses.
2206 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2207 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2208 AddUsersIfInteresting(I, L, Processed);
2210 if (!IVUsesByStride.empty()) {
2211 // Optimize induction variables. Some indvar uses can be transformed to use
2212 // strides that will be needed for other purposes. A common example of this
2213 // is the exit test for the loop, which can often be rewritten to use the
2214 // computation of some other indvar to decide when to terminate the loop.
2217 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2218 // doing computation in byte values, promote to 32-bit values if safe.
2220 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2221 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2222 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2223 // Need to be careful that IV's are all the same type. Only works for
2224 // intptr_t indvars.
2226 // If we only have one stride, we can more aggressively eliminate some
2228 bool HasOneStride = IVUsesByStride.size() == 1;
2231 DOUT << "\nLSR on ";
2235 // IVsByStride keeps IVs for one particular loop.
2236 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2238 // Sort the StrideOrder so we process larger strides first.
2239 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2241 // Note: this processes each stride/type pair individually. All users
2242 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2243 // Also, note that we iterate over IVUsesByStride indirectly by using
2244 // StrideOrder. This extra layer of indirection makes the ordering of
2245 // strides deterministic - not dependent on map order.
2246 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2247 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2248 IVUsesByStride.find(StrideOrder[Stride]);
2249 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2250 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2254 // We're done analyzing this loop; release all the state we built up for it.
2255 CastedPointers.clear();
2256 IVUsesByStride.clear();
2257 IVsByStride.clear();
2258 StrideOrder.clear();
2259 for (unsigned i=0; i<GEPlist.size(); i++)
2260 SE->deleteValueFromRecords(GEPlist[i]);
2263 // Clean up after ourselves
2264 if (!DeadInsts.empty()) {
2265 DeleteTriviallyDeadInstructions();
2267 BasicBlock::iterator I = L->getHeader()->begin();
2268 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2269 // At this point, we know that we have killed one or more IV users.
2270 // It is worth checking to see if the cannonical indvar is also
2271 // dead, so that we can remove it as well.
2273 // We can remove a PHI if it is on a cycle in the def-use graph
2274 // where each node in the cycle has degree one, i.e. only one use,
2275 // and is an instruction with no side effects.
2277 // FIXME: this needs to eliminate an induction variable even if it's being
2278 // compared against some value to decide loop termination.
2279 if (!PN->hasOneUse())
2282 SmallPtrSet<PHINode *, 4> PHIs;
2283 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2284 J && J->hasOneUse() && !J->mayWriteToMemory();
2285 J = dyn_cast<Instruction>(*J->use_begin())) {
2286 // If we find the original PHI, we've discovered a cycle.
2288 // Break the cycle and mark the PHI for deletion.
2289 SE->deleteValueFromRecords(PN);
2290 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2291 DeadInsts.push_back(PN);
2295 // If we find a PHI more than once, we're on a cycle that
2296 // won't prove fruitful.
2297 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2301 DeleteTriviallyDeadInstructions();