1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs 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.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "loop-reduce"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/Type.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/ScalarEvolutionExpander.h"
26 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/Compiler.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Target/TargetLowering.h"
40 STATISTIC(NumReduced , "Number of IV uses strength reduced");
41 STATISTIC(NumInserted, "Number of PHIs inserted");
42 STATISTIC(NumVariable, "Number of PHIs with variable strides");
43 STATISTIC(NumEliminated, "Number of strides eliminated");
44 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
45 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
47 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
55 /// IVStrideUse - Keep track of one use of a strided induction variable, where
56 /// the stride is stored externally. The Offset member keeps track of the
57 /// offset from the IV, User is the actual user of the operand, and
58 /// 'OperandValToReplace' is the operand of the User that is the use.
59 struct VISIBILITY_HIDDEN IVStrideUse {
62 Value *OperandValToReplace;
64 // isUseOfPostIncrementedValue - True if this should use the
65 // post-incremented version of this IV, not the preincremented version.
66 // This can only be set in special cases, such as the terminating setcc
67 // instruction for a loop or uses dominated by the loop.
68 bool isUseOfPostIncrementedValue;
70 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
71 : Offset(Offs), User(U), OperandValToReplace(O),
72 isUseOfPostIncrementedValue(false) {}
75 /// IVUsersOfOneStride - This structure keeps track of all instructions that
76 /// have an operand that is based on the trip count multiplied by some stride.
77 /// The stride for all of these users is common and kept external to this
79 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
80 /// Users - Keep track of all of the users of this stride as well as the
81 /// initial value and the operand that uses the IV.
82 std::vector<IVStrideUse> Users;
84 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
85 Users.push_back(IVStrideUse(Offset, User, Operand));
89 /// IVInfo - This structure keeps track of one IV expression inserted during
90 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
91 /// well as the PHI node and increment value created for rewrite.
92 struct VISIBILITY_HIDDEN IVExpr {
97 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi)
98 : Stride(stride), Base(base), PHI(phi) {}
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) {
107 IVs.push_back(IVExpr(Stride, Base, PHI));
111 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
115 const TargetData *TD;
116 const Type *UIntPtrTy;
119 /// IVUsesByStride - Keep track of all uses of induction variables that we
120 /// are interested in. The key of the map is the stride of the access.
121 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
123 /// IVsByStride - Keep track of all IVs that have been inserted for a
124 /// particular stride.
125 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
127 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
128 /// We use this to iterate over the IVUsesByStride collection without being
129 /// dependent on random ordering of pointers in the process.
130 SmallVector<SCEVHandle, 16> StrideOrder;
132 /// DeadInsts - Keep track of instructions we may have made dead, so that
133 /// we can remove them after we are done working.
134 SmallVector<Instruction*, 16> DeadInsts;
136 /// TLI - Keep a pointer of a TargetLowering to consult for determining
137 /// transformation profitability.
138 const TargetLowering *TLI;
141 static char ID; // Pass ID, replacement for typeid
142 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
143 LoopPass(&ID), TLI(tli) {
146 bool runOnLoop(Loop *L, LPPassManager &LPM);
148 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
149 // We split critical edges, so we change the CFG. However, we do update
150 // many analyses if they are around.
151 AU.addPreservedID(LoopSimplifyID);
152 AU.addPreserved<LoopInfo>();
153 AU.addPreserved<DominanceFrontier>();
154 AU.addPreserved<DominatorTree>();
156 AU.addRequiredID(LoopSimplifyID);
157 AU.addRequired<LoopInfo>();
158 AU.addRequired<DominatorTree>();
159 AU.addRequired<TargetData>();
160 AU.addRequired<ScalarEvolution>();
161 AU.addPreserved<ScalarEvolution>();
165 bool AddUsersIfInteresting(Instruction *I, Loop *L,
166 SmallPtrSet<Instruction*,16> &Processed);
167 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
168 IVStrideUse* &CondUse,
169 const SCEVHandle* &CondStride);
170 void OptimizeIndvars(Loop *L);
172 /// OptimizeShadowIV - If IV is used in a int-to-float cast
173 /// inside the loop then try to eliminate the cast opeation.
174 void OptimizeShadowIV(Loop *L);
176 /// OptimizeSMax - Rewrite the loop's terminating condition
177 /// if it uses an smax computation.
178 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
179 IVStrideUse* &CondUse);
181 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
182 const SCEVHandle *&CondStride);
183 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
184 SCEVHandle CheckForIVReuse(bool, bool, bool, const SCEVHandle&,
185 IVExpr&, const Type*,
186 const std::vector<BasedUser>& UsersToProcess);
187 bool ValidStride(bool, int64_t,
188 const std::vector<BasedUser>& UsersToProcess);
189 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
190 IVUsersOfOneStride &Uses,
192 bool &AllUsesAreAddresses,
193 bool &AllUsesAreOutsideLoop,
194 std::vector<BasedUser> &UsersToProcess);
195 bool ShouldUseFullStrengthReductionMode(
196 const std::vector<BasedUser> &UsersToProcess,
198 bool AllUsesAreAddresses,
200 void PrepareToStrengthReduceFully(
201 std::vector<BasedUser> &UsersToProcess,
203 SCEVHandle CommonExprs,
205 SCEVExpander &PreheaderRewriter);
206 void PrepareToStrengthReduceFromSmallerStride(
207 std::vector<BasedUser> &UsersToProcess,
209 const IVExpr &ReuseIV,
210 Instruction *PreInsertPt);
211 void PrepareToStrengthReduceWithNewPhi(
212 std::vector<BasedUser> &UsersToProcess,
214 SCEVHandle CommonExprs,
217 SCEVExpander &PreheaderRewriter);
218 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
219 IVUsersOfOneStride &Uses,
221 void DeleteTriviallyDeadInstructions();
225 char LoopStrengthReduce::ID = 0;
226 static RegisterPass<LoopStrengthReduce>
227 X("loop-reduce", "Loop Strength Reduction");
229 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
230 return new LoopStrengthReduce(TLI);
233 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
234 /// specified set are trivially dead, delete them and see if this makes any of
235 /// their operands subsequently dead.
236 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
237 if (DeadInsts.empty()) return;
239 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
240 // go. The code below never adds a non-dead instruction to the worklist, but
241 // callers may not be so careful.
242 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
244 // Drop duplicate instructions and those with uses.
245 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
246 Instruction *I = DeadInsts[i];
247 if (!I->use_empty()) DeadInsts[i] = 0;
248 while (i != e && DeadInsts[i+1] == I)
252 while (!DeadInsts.empty()) {
253 Instruction *I = DeadInsts.back();
254 DeadInsts.pop_back();
256 if (I == 0 || !isInstructionTriviallyDead(I))
259 SE->deleteValueFromRecords(I);
261 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
262 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
265 DeadInsts.push_back(U);
269 I->eraseFromParent();
274 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
275 /// subexpression that is an AddRec from a loop other than L. An outer loop
276 /// of L is OK, but not an inner loop nor a disjoint loop.
277 static bool containsAddRecFromDifferentLoop(SCEVHandle S, Loop *L) {
278 // This is very common, put it first.
279 if (isa<SCEVConstant>(S))
281 if (SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
282 for (unsigned int i=0; i< AE->getNumOperands(); i++)
283 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
287 if (SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
288 if (const Loop *newLoop = AE->getLoop()) {
291 // if newLoop is an outer loop of L, this is OK.
292 if (!LoopInfoBase<BasicBlock>::isNotAlreadyContainedIn(L, newLoop))
297 if (SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
298 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
299 containsAddRecFromDifferentLoop(DE->getRHS(), L);
301 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
302 // need this when it is.
303 if (SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
304 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
305 containsAddRecFromDifferentLoop(DE->getRHS(), L);
307 if (SCEVTruncateExpr *TE = dyn_cast<SCEVTruncateExpr>(S))
308 return containsAddRecFromDifferentLoop(TE->getOperand(), L);
309 if (SCEVZeroExtendExpr *ZE = dyn_cast<SCEVZeroExtendExpr>(S))
310 return containsAddRecFromDifferentLoop(ZE->getOperand(), L);
311 if (SCEVSignExtendExpr *SE = dyn_cast<SCEVSignExtendExpr>(S))
312 return containsAddRecFromDifferentLoop(SE->getOperand(), L);
316 /// getSCEVStartAndStride - Compute the start and stride of this expression,
317 /// returning false if the expression is not a start/stride pair, or true if it
318 /// is. The stride must be a loop invariant expression, but the start may be
319 /// a mix of loop invariant and loop variant expressions. The start cannot,
320 /// however, contain an AddRec from a different loop, unless that loop is an
321 /// outer loop of the current loop.
322 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
323 SCEVHandle &Start, SCEVHandle &Stride,
324 ScalarEvolution *SE, DominatorTree *DT) {
325 SCEVHandle TheAddRec = Start; // Initialize to zero.
327 // If the outer level is an AddExpr, the operands are all start values except
328 // for a nested AddRecExpr.
329 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
330 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
331 if (SCEVAddRecExpr *AddRec =
332 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
333 if (AddRec->getLoop() == L)
334 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
336 return false; // Nested IV of some sort?
338 Start = SE->getAddExpr(Start, AE->getOperand(i));
341 } else if (isa<SCEVAddRecExpr>(SH)) {
344 return false; // not analyzable.
347 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
348 if (!AddRec || AddRec->getLoop() != L) return false;
350 // FIXME: Generalize to non-affine IV's.
351 if (!AddRec->isAffine()) return false;
353 // If Start contains an SCEVAddRecExpr from a different loop, other than an
354 // outer loop of the current loop, reject it. SCEV has no concept of
355 // operating on more than one loop at a time so don't confuse it with such
357 if (containsAddRecFromDifferentLoop(AddRec->getOperand(0), L))
360 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
362 if (!isa<SCEVConstant>(AddRec->getOperand(1))) {
363 // If stride is an instruction, make sure it dominates the loop preheader.
364 // Otherwise we could end up with a use before def situation.
365 BasicBlock *Preheader = L->getLoopPreheader();
366 if (!AddRec->getOperand(1)->dominates(Preheader, DT))
369 DOUT << "[" << L->getHeader()->getName()
370 << "] Variable stride: " << *AddRec << "\n";
373 Stride = AddRec->getOperand(1);
377 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
378 /// and now we need to decide whether the user should use the preinc or post-inc
379 /// value. If this user should use the post-inc version of the IV, return true.
381 /// Choosing wrong here can break dominance properties (if we choose to use the
382 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
383 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
384 /// should use the post-inc value).
385 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
386 Loop *L, DominatorTree *DT, Pass *P,
387 SmallVectorImpl<Instruction*> &DeadInsts){
388 // If the user is in the loop, use the preinc value.
389 if (L->contains(User->getParent())) return false;
391 BasicBlock *LatchBlock = L->getLoopLatch();
393 // Ok, the user is outside of the loop. If it is dominated by the latch
394 // block, use the post-inc value.
395 if (DT->dominates(LatchBlock, User->getParent()))
398 // There is one case we have to be careful of: PHI nodes. These little guys
399 // can live in blocks that do not dominate the latch block, but (since their
400 // uses occur in the predecessor block, not the block the PHI lives in) should
401 // still use the post-inc value. Check for this case now.
402 PHINode *PN = dyn_cast<PHINode>(User);
403 if (!PN) return false; // not a phi, not dominated by latch block.
405 // Look at all of the uses of IV by the PHI node. If any use corresponds to
406 // a block that is not dominated by the latch block, give up and use the
407 // preincremented value.
408 unsigned NumUses = 0;
409 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
410 if (PN->getIncomingValue(i) == IV) {
412 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
416 // Okay, all uses of IV by PN are in predecessor blocks that really are
417 // dominated by the latch block. Split the critical edges and use the
418 // post-incremented value.
419 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
420 if (PN->getIncomingValue(i) == IV) {
421 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
422 // Splitting the critical edge can reduce the number of entries in this
424 e = PN->getNumIncomingValues();
425 if (--NumUses == 0) break;
428 // PHI node might have become a constant value after SplitCriticalEdge.
429 DeadInsts.push_back(User);
434 /// isAddressUse - Returns true if the specified instruction is using the
435 /// specified value as an address.
436 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
437 bool isAddress = isa<LoadInst>(Inst);
438 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
439 if (SI->getOperand(1) == OperandVal)
441 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
442 // Addressing modes can also be folded into prefetches and a variety
444 switch (II->getIntrinsicID()) {
446 case Intrinsic::prefetch:
447 case Intrinsic::x86_sse2_loadu_dq:
448 case Intrinsic::x86_sse2_loadu_pd:
449 case Intrinsic::x86_sse_loadu_ps:
450 case Intrinsic::x86_sse_storeu_ps:
451 case Intrinsic::x86_sse2_storeu_pd:
452 case Intrinsic::x86_sse2_storeu_dq:
453 case Intrinsic::x86_sse2_storel_dq:
454 if (II->getOperand(1) == OperandVal)
462 /// getAccessType - Return the type of the memory being accessed.
463 static const Type *getAccessType(const Instruction *Inst) {
464 const Type *UseTy = Inst->getType();
465 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
466 UseTy = SI->getOperand(0)->getType();
467 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
468 // Addressing modes can also be folded into prefetches and a variety
470 switch (II->getIntrinsicID()) {
472 case Intrinsic::x86_sse_storeu_ps:
473 case Intrinsic::x86_sse2_storeu_pd:
474 case Intrinsic::x86_sse2_storeu_dq:
475 case Intrinsic::x86_sse2_storel_dq:
476 UseTy = II->getOperand(1)->getType();
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.
491 // LSR is not APInt clean, do not touch integers bigger than 64-bits.
492 if (I->getType()->isInteger() &&
493 I->getType()->getPrimitiveSizeInBits() > 64)
496 if (!Processed.insert(I))
497 return true; // Instruction already handled.
499 // Get the symbolic expression for this instruction.
500 SCEVHandle ISE = SE->getSCEV(I);
501 if (isa<SCEVCouldNotCompute>(ISE)) return false;
503 // Get the start and stride for this expression.
504 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
505 SCEVHandle Stride = Start;
506 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE, DT))
507 return false; // Non-reducible symbolic expression, bail out.
509 std::vector<Instruction *> IUsers;
510 // Collect all I uses now because IVUseShouldUsePostIncValue may
511 // invalidate use_iterator.
512 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
513 IUsers.push_back(cast<Instruction>(*UI));
515 for (unsigned iused_index = 0, iused_size = IUsers.size();
516 iused_index != iused_size; ++iused_index) {
518 Instruction *User = IUsers[iused_index];
520 // Do not infinitely recurse on PHI nodes.
521 if (isa<PHINode>(User) && Processed.count(User))
524 // Descend recursively, but not into PHI nodes outside the current loop.
525 // It's important to see the entire expression outside the loop to get
526 // choices that depend on addressing mode use right, although we won't
527 // consider references ouside the loop in all cases.
528 // If User is already in Processed, we don't want to recurse into it again,
529 // but do want to record a second reference in the same instruction.
530 bool AddUserToIVUsers = false;
531 if (LI->getLoopFor(User->getParent()) != L) {
532 if (isa<PHINode>(User) || Processed.count(User) ||
533 !AddUsersIfInteresting(User, L, Processed)) {
534 DOUT << "FOUND USER in other loop: " << *User
535 << " OF SCEV: " << *ISE << "\n";
536 AddUserToIVUsers = true;
538 } else if (Processed.count(User) ||
539 !AddUsersIfInteresting(User, L, Processed)) {
540 DOUT << "FOUND USER: " << *User
541 << " OF SCEV: " << *ISE << "\n";
542 AddUserToIVUsers = true;
545 if (AddUserToIVUsers) {
546 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
547 if (StrideUses.Users.empty()) // First occurrence of this stride?
548 StrideOrder.push_back(Stride);
550 // Okay, we found a user that we cannot reduce. Analyze the instruction
551 // and decide what to do with it. If we are a use inside of the loop, use
552 // the value before incrementation, otherwise use it after incrementation.
553 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
554 // The value used will be incremented by the stride more than we are
555 // expecting, so subtract this off.
556 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
557 StrideUses.addUser(NewStart, User, I);
558 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
559 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
561 StrideUses.addUser(Start, User, I);
569 /// BasedUser - For a particular base value, keep information about how we've
570 /// partitioned the expression so far.
572 /// SE - The current ScalarEvolution object.
575 /// Base - The Base value for the PHI node that needs to be inserted for
576 /// this use. As the use is processed, information gets moved from this
577 /// field to the Imm field (below). BasedUser values are sorted by this
581 /// Inst - The instruction using the induction variable.
584 /// OperandValToReplace - The operand value of Inst to replace with the
586 Value *OperandValToReplace;
588 /// Imm - The immediate value that should be added to the base immediately
589 /// before Inst, because it will be folded into the imm field of the
590 /// instruction. This is also sometimes used for loop-variant values that
591 /// must be added inside the loop.
594 /// Phi - The induction variable that performs the striding that
595 /// should be used for this user.
598 // isUseOfPostIncrementedValue - True if this should use the
599 // post-incremented version of this IV, not the preincremented version.
600 // This can only be set in special cases, such as the terminating setcc
601 // instruction for a loop and uses outside the loop that are dominated by
603 bool isUseOfPostIncrementedValue;
605 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
606 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
607 OperandValToReplace(IVSU.OperandValToReplace),
608 Imm(SE->getIntegerSCEV(0, Base->getType())),
609 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
611 // Once we rewrite the code to insert the new IVs we want, update the
612 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
614 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
615 Instruction *InsertPt,
616 SCEVExpander &Rewriter, Loop *L, Pass *P,
617 SmallVectorImpl<Instruction*> &DeadInsts);
619 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
621 SCEVExpander &Rewriter,
622 Instruction *IP, Loop *L);
627 void BasedUser::dump() const {
628 cerr << " Base=" << *Base;
629 cerr << " Imm=" << *Imm;
630 cerr << " Inst: " << *Inst;
633 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
635 SCEVExpander &Rewriter,
636 Instruction *IP, Loop *L) {
637 // Figure out where we *really* want to insert this code. In particular, if
638 // the user is inside of a loop that is nested inside of L, we really don't
639 // want to insert this expression before the user, we'd rather pull it out as
640 // many loops as possible.
641 LoopInfo &LI = Rewriter.getLoopInfo();
642 Instruction *BaseInsertPt = IP;
644 // Figure out the most-nested loop that IP is in.
645 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
647 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
648 // the preheader of the outer-most loop where NewBase is not loop invariant.
649 if (L->contains(IP->getParent()))
650 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
651 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
652 InsertLoop = InsertLoop->getParentLoop();
655 Value *Base = Rewriter.expandCodeFor(NewBase, Ty, BaseInsertPt);
657 // If there is no immediate value, skip the next part.
661 // If we are inserting the base and imm values in the same block, make sure to
662 // adjust the IP position if insertion reused a result.
663 if (IP == BaseInsertPt)
664 IP = Rewriter.getInsertionPoint();
666 // Always emit the immediate (if non-zero) into the same block as the user.
667 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
668 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
672 // Once we rewrite the code to insert the new IVs we want, update the
673 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
674 // to it. NewBasePt is the last instruction which contributes to the
675 // value of NewBase in the case that it's a diffferent instruction from
676 // the PHI that NewBase is computed from, or null otherwise.
678 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
679 Instruction *NewBasePt,
680 SCEVExpander &Rewriter, Loop *L, Pass *P,
681 SmallVectorImpl<Instruction*> &DeadInsts){
682 if (!isa<PHINode>(Inst)) {
683 // By default, insert code at the user instruction.
684 BasicBlock::iterator InsertPt = Inst;
686 // However, if the Operand is itself an instruction, the (potentially
687 // complex) inserted code may be shared by many users. Because of this, we
688 // want to emit code for the computation of the operand right before its old
689 // computation. This is usually safe, because we obviously used to use the
690 // computation when it was computed in its current block. However, in some
691 // cases (e.g. use of a post-incremented induction variable) the NewBase
692 // value will be pinned to live somewhere after the original computation.
693 // In this case, we have to back off.
695 // If this is a use outside the loop (which means after, since it is based
696 // on a loop indvar) we use the post-incremented value, so that we don't
697 // artificially make the preinc value live out the bottom of the loop.
698 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
699 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
700 InsertPt = NewBasePt;
702 } else if (Instruction *OpInst
703 = dyn_cast<Instruction>(OperandValToReplace)) {
705 while (isa<PHINode>(InsertPt)) ++InsertPt;
708 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
709 OperandValToReplace->getType(),
710 Rewriter, InsertPt, L);
711 // Replace the use of the operand Value with the new Phi we just created.
712 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
714 DOUT << " Replacing with ";
715 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
716 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
720 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
721 // expression into each operand block that uses it. Note that PHI nodes can
722 // have multiple entries for the same predecessor. We use a map to make sure
723 // that a PHI node only has a single Value* for each predecessor (which also
724 // prevents us from inserting duplicate code in some blocks).
725 DenseMap<BasicBlock*, Value*> InsertedCode;
726 PHINode *PN = cast<PHINode>(Inst);
727 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
728 if (PN->getIncomingValue(i) == OperandValToReplace) {
729 // If the original expression is outside the loop, put the replacement
730 // code in the same place as the original expression,
731 // which need not be an immediate predecessor of this PHI. This way we
732 // need only one copy of it even if it is referenced multiple times in
733 // the PHI. We don't do this when the original expression is inside the
734 // loop because multiple copies sometimes do useful sinking of code in
736 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
737 if (L->contains(OldLoc->getParent())) {
738 // If this is a critical edge, split the edge so that we do not insert
739 // the code on all predecessor/successor paths. We do this unless this
740 // is the canonical backedge for this loop, as this can make some
741 // inserted code be in an illegal position.
742 BasicBlock *PHIPred = PN->getIncomingBlock(i);
743 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
744 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
746 // First step, split the critical edge.
747 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
749 // Next step: move the basic block. In particular, if the PHI node
750 // is outside of the loop, and PredTI is in the loop, we want to
751 // move the block to be immediately before the PHI block, not
752 // immediately after PredTI.
753 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
754 BasicBlock *NewBB = PN->getIncomingBlock(i);
755 NewBB->moveBefore(PN->getParent());
758 // Splitting the edge can reduce the number of PHI entries we have.
759 e = PN->getNumIncomingValues();
762 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
764 // Insert the code into the end of the predecessor block.
765 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
766 PN->getIncomingBlock(i)->getTerminator() :
767 OldLoc->getParent()->getTerminator();
768 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
769 Rewriter, InsertPt, L);
771 DOUT << " Changing PHI use to ";
772 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
773 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
776 // Replace the use of the operand Value with the new Phi we just created.
777 PN->setIncomingValue(i, Code);
782 // PHI node might have become a constant value after SplitCriticalEdge.
783 DeadInsts.push_back(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 (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
806 TargetLowering::AddrMode AM;
808 AM.HasBaseReg = HasBaseReg;
809 return TLI->isLegalAddressingMode(AM, UseTy);
815 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
816 /// loop varying to the Imm operand.
817 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
818 Loop *L, ScalarEvolution *SE) {
819 if (Val->isLoopInvariant(L)) return; // Nothing to do.
821 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
822 std::vector<SCEVHandle> NewOps;
823 NewOps.reserve(SAE->getNumOperands());
825 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
826 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
827 // If this is a loop-variant expression, it must stay in the immediate
828 // field of the expression.
829 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
831 NewOps.push_back(SAE->getOperand(i));
835 Val = SE->getIntegerSCEV(0, Val->getType());
837 Val = SE->getAddExpr(NewOps);
838 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
839 // Try to pull immediates out of the start value of nested addrec's.
840 SCEVHandle Start = SARE->getStart();
841 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
843 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
845 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
847 // Otherwise, all of Val is variant, move the whole thing over.
848 Imm = SE->getAddExpr(Imm, Val);
849 Val = SE->getIntegerSCEV(0, Val->getType());
854 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
855 /// that can fit into the immediate field of instructions in the target.
856 /// Accumulate these immediate values into the Imm value.
857 static void MoveImmediateValues(const TargetLowering *TLI,
859 SCEVHandle &Val, SCEVHandle &Imm,
860 bool isAddress, Loop *L,
861 ScalarEvolution *SE) {
862 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
863 std::vector<SCEVHandle> NewOps;
864 NewOps.reserve(SAE->getNumOperands());
866 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
867 SCEVHandle NewOp = SAE->getOperand(i);
868 MoveImmediateValues(TLI, UseTy, NewOp, Imm, isAddress, L, SE);
870 if (!NewOp->isLoopInvariant(L)) {
871 // If this is a loop-variant expression, it must stay in the immediate
872 // field of the expression.
873 Imm = SE->getAddExpr(Imm, NewOp);
875 NewOps.push_back(NewOp);
880 Val = SE->getIntegerSCEV(0, Val->getType());
882 Val = SE->getAddExpr(NewOps);
884 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
885 // Try to pull immediates out of the start value of nested addrec's.
886 SCEVHandle Start = SARE->getStart();
887 MoveImmediateValues(TLI, UseTy, Start, Imm, isAddress, L, SE);
889 if (Start != SARE->getStart()) {
890 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
892 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
895 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
896 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
897 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
898 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
900 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
901 SCEVHandle NewOp = SME->getOperand(1);
902 MoveImmediateValues(TLI, UseTy, NewOp, SubImm, isAddress, L, SE);
904 // If we extracted something out of the subexpressions, see if we can
906 if (NewOp != SME->getOperand(1)) {
907 // Scale SubImm up by "8". If the result is a target constant, we are
909 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
910 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
911 // Accumulate the immediate.
912 Imm = SE->getAddExpr(Imm, SubImm);
914 // Update what is left of 'Val'.
915 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
922 // Loop-variant expressions must stay in the immediate field of the
924 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
925 !Val->isLoopInvariant(L)) {
926 Imm = SE->getAddExpr(Imm, Val);
927 Val = SE->getIntegerSCEV(0, Val->getType());
931 // Otherwise, no immediates to move.
934 static void MoveImmediateValues(const TargetLowering *TLI,
936 SCEVHandle &Val, SCEVHandle &Imm,
937 bool isAddress, Loop *L,
938 ScalarEvolution *SE) {
939 const Type *UseTy = getAccessType(User);
940 MoveImmediateValues(TLI, UseTy, Val, Imm, isAddress, L, SE);
943 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
944 /// added together. This is used to reassociate common addition subexprs
945 /// together for maximal sharing when rewriting bases.
946 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
948 ScalarEvolution *SE) {
949 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
950 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
951 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
952 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
953 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
954 if (SARE->getOperand(0) == Zero) {
955 SubExprs.push_back(Expr);
957 // Compute the addrec with zero as its base.
958 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
959 Ops[0] = Zero; // Start with zero base.
960 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
963 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
965 } else if (!Expr->isZero()) {
967 SubExprs.push_back(Expr);
971 // This is logically local to the following function, but C++ says we have
972 // to make it file scope.
973 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
975 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
976 /// the Uses, removing any common subexpressions, except that if all such
977 /// subexpressions can be folded into an addressing mode for all uses inside
978 /// the loop (this case is referred to as "free" in comments herein) we do
979 /// not remove anything. This looks for things like (a+b+c) and
980 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
981 /// is *removed* from the Bases and returned.
983 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
984 ScalarEvolution *SE, Loop *L,
985 const TargetLowering *TLI) {
986 unsigned NumUses = Uses.size();
988 // Only one use? This is a very common case, so we handle it specially and
990 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
991 SCEVHandle Result = Zero;
992 SCEVHandle FreeResult = Zero;
994 // If the use is inside the loop, use its base, regardless of what it is:
995 // it is clearly shared across all the IV's. If the use is outside the loop
996 // (which means after it) we don't want to factor anything *into* the loop,
997 // so just use 0 as the base.
998 if (L->contains(Uses[0].Inst->getParent()))
999 std::swap(Result, Uses[0].Base);
1003 // To find common subexpressions, count how many of Uses use each expression.
1004 // If any subexpressions are used Uses.size() times, they are common.
1005 // Also track whether all uses of each expression can be moved into an
1006 // an addressing mode "for free"; such expressions are left within the loop.
1007 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1008 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
1010 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
1011 // order we see them.
1012 std::vector<SCEVHandle> UniqueSubExprs;
1014 std::vector<SCEVHandle> SubExprs;
1015 unsigned NumUsesInsideLoop = 0;
1016 for (unsigned i = 0; i != NumUses; ++i) {
1017 // If the user is outside the loop, just ignore it for base computation.
1018 // Since the user is outside the loop, it must be *after* the loop (if it
1019 // were before, it could not be based on the loop IV). We don't want users
1020 // after the loop to affect base computation of values *inside* the loop,
1021 // because we can always add their offsets to the result IV after the loop
1022 // is done, ensuring we get good code inside the loop.
1023 if (!L->contains(Uses[i].Inst->getParent()))
1025 NumUsesInsideLoop++;
1027 // If the base is zero (which is common), return zero now, there are no
1028 // CSEs we can find.
1029 if (Uses[i].Base == Zero) return Zero;
1031 // If this use is as an address we may be able to put CSEs in the addressing
1032 // mode rather than hoisting them.
1033 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1034 // We may need the UseTy below, but only when isAddrUse, so compute it
1035 // only in that case.
1036 const Type *UseTy = 0;
1038 UseTy = getAccessType(Uses[i].Inst);
1040 // Split the expression into subexprs.
1041 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1042 // Add one to SubExpressionUseData.Count for each subexpr present, and
1043 // if the subexpr is not a valid immediate within an addressing mode use,
1044 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1045 // hoist these out of the loop (if they are common to all uses).
1046 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1047 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1048 UniqueSubExprs.push_back(SubExprs[j]);
1049 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1050 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1055 // Now that we know how many times each is used, build Result. Iterate over
1056 // UniqueSubexprs so that we have a stable ordering.
1057 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1058 std::map<SCEVHandle, SubExprUseData>::iterator I =
1059 SubExpressionUseData.find(UniqueSubExprs[i]);
1060 assert(I != SubExpressionUseData.end() && "Entry not found?");
1061 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1062 if (I->second.notAllUsesAreFree)
1063 Result = SE->getAddExpr(Result, I->first);
1065 FreeResult = SE->getAddExpr(FreeResult, I->first);
1067 // Remove non-cse's from SubExpressionUseData.
1068 SubExpressionUseData.erase(I);
1071 if (FreeResult != Zero) {
1072 // We have some subexpressions that can be subsumed into addressing
1073 // modes in every use inside the loop. However, it's possible that
1074 // there are so many of them that the combined FreeResult cannot
1075 // be subsumed, or that the target cannot handle both a FreeResult
1076 // and a Result in the same instruction (for example because it would
1077 // require too many registers). Check this.
1078 for (unsigned i=0; i<NumUses; ++i) {
1079 if (!L->contains(Uses[i].Inst->getParent()))
1081 // We know this is an addressing mode use; if there are any uses that
1082 // are not, FreeResult would be Zero.
1083 const Type *UseTy = getAccessType(Uses[i].Inst);
1084 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1085 // FIXME: could split up FreeResult into pieces here, some hoisted
1086 // and some not. There is no obvious advantage to this.
1087 Result = SE->getAddExpr(Result, FreeResult);
1094 // If we found no CSE's, return now.
1095 if (Result == Zero) return Result;
1097 // If we still have a FreeResult, remove its subexpressions from
1098 // SubExpressionUseData. This means they will remain in the use Bases.
1099 if (FreeResult != Zero) {
1100 SeparateSubExprs(SubExprs, FreeResult, SE);
1101 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1102 std::map<SCEVHandle, SubExprUseData>::iterator I =
1103 SubExpressionUseData.find(SubExprs[j]);
1104 SubExpressionUseData.erase(I);
1109 // Otherwise, remove all of the CSE's we found from each of the base values.
1110 for (unsigned i = 0; i != NumUses; ++i) {
1111 // Uses outside the loop don't necessarily include the common base, but
1112 // the final IV value coming into those uses does. Instead of trying to
1113 // remove the pieces of the common base, which might not be there,
1114 // subtract off the base to compensate for this.
1115 if (!L->contains(Uses[i].Inst->getParent())) {
1116 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1120 // Split the expression into subexprs.
1121 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1123 // Remove any common subexpressions.
1124 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1125 if (SubExpressionUseData.count(SubExprs[j])) {
1126 SubExprs.erase(SubExprs.begin()+j);
1130 // Finally, add the non-shared expressions together.
1131 if (SubExprs.empty())
1132 Uses[i].Base = Zero;
1134 Uses[i].Base = SE->getAddExpr(SubExprs);
1141 /// ValidStride - Check whether the given Scale is valid for all loads and
1142 /// stores in UsersToProcess.
1144 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1146 const std::vector<BasedUser>& UsersToProcess) {
1150 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1151 // If this is a load or other access, pass the type of the access in.
1152 const Type *AccessTy = Type::VoidTy;
1153 if (isAddressUse(UsersToProcess[i].Inst,
1154 UsersToProcess[i].OperandValToReplace))
1155 AccessTy = getAccessType(UsersToProcess[i].Inst);
1156 else if (isa<PHINode>(UsersToProcess[i].Inst))
1159 TargetLowering::AddrMode AM;
1160 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1161 AM.BaseOffs = SC->getValue()->getSExtValue();
1162 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1165 // If load[imm+r*scale] is illegal, bail out.
1166 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1172 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
1174 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1178 if (Ty1->canLosslesslyBitCastTo(Ty2))
1180 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1182 if (isa<PointerType>(Ty2) && Ty1->canLosslesslyBitCastTo(UIntPtrTy))
1184 if (isa<PointerType>(Ty1) && Ty2->canLosslesslyBitCastTo(UIntPtrTy))
1189 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1190 /// of a previous stride and it is a legal value for the target addressing
1191 /// mode scale component and optional base reg. This allows the users of
1192 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1193 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1195 /// If all uses are outside the loop, we don't require that all multiplies
1196 /// be folded into the addressing mode, nor even that the factor be constant;
1197 /// a multiply (executed once) outside the loop is better than another IV
1198 /// within. Well, usually.
1199 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1200 bool AllUsesAreAddresses,
1201 bool AllUsesAreOutsideLoop,
1202 const SCEVHandle &Stride,
1203 IVExpr &IV, const Type *Ty,
1204 const std::vector<BasedUser>& UsersToProcess) {
1205 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1206 int64_t SInt = SC->getValue()->getSExtValue();
1207 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1209 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1210 IVsByStride.find(StrideOrder[NewStride]);
1211 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1213 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1214 if (SI->first != Stride &&
1215 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1217 int64_t Scale = SInt / SSInt;
1218 // Check that this stride is valid for all the types used for loads and
1219 // stores; if it can be used for some and not others, we might as well use
1220 // the original stride everywhere, since we have to create the IV for it
1221 // anyway. If the scale is 1, then we don't need to worry about folding
1224 (AllUsesAreAddresses &&
1225 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1226 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1227 IE = SI->second.IVs.end(); II != IE; ++II)
1228 // FIXME: Only handle base == 0 for now.
1229 // Only reuse previous IV if it would not require a type conversion.
1230 if (II->Base->isZero() &&
1231 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1233 return SE->getIntegerSCEV(Scale, Stride->getType());
1236 } else if (AllUsesAreOutsideLoop) {
1237 // Accept nonconstant strides here; it is really really right to substitute
1238 // an existing IV if we can.
1239 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1241 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1242 IVsByStride.find(StrideOrder[NewStride]);
1243 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1245 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1246 if (SI->first != Stride && SSInt != 1)
1248 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1249 IE = SI->second.IVs.end(); II != IE; ++II)
1250 // Accept nonzero base here.
1251 // Only reuse previous IV if it would not require a type conversion.
1252 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1257 // Special case, old IV is -1*x and this one is x. Can treat this one as
1259 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1261 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1262 IVsByStride.find(StrideOrder[NewStride]);
1263 if (SI == IVsByStride.end())
1265 if (SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1266 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1267 if (Stride == ME->getOperand(1) &&
1268 SC->getValue()->getSExtValue() == -1LL)
1269 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1270 IE = SI->second.IVs.end(); II != IE; ++II)
1271 // Accept nonzero base here.
1272 // Only reuse previous IV if it would not require type conversion.
1273 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1275 return SE->getIntegerSCEV(-1LL, Stride->getType());
1279 return SE->getIntegerSCEV(0, Stride->getType());
1282 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1283 /// returns true if Val's isUseOfPostIncrementedValue is true.
1284 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1285 return Val.isUseOfPostIncrementedValue;
1288 /// isNonConstantNegative - Return true if the specified scev is negated, but
1290 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1291 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1292 if (!Mul) return false;
1294 // If there is a constant factor, it will be first.
1295 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1296 if (!SC) return false;
1298 // Return true if the value is negative, this matches things like (-42 * V).
1299 return SC->getValue()->getValue().isNegative();
1302 // CollectIVUsers - Transform our list of users and offsets to a bit more
1303 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1304 // of the strided accesses, as well as the old information from Uses. We
1305 // progressively move information from the Base field to the Imm field, until
1306 // we eventually have the full access expression to rewrite the use.
1307 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1308 IVUsersOfOneStride &Uses,
1310 bool &AllUsesAreAddresses,
1311 bool &AllUsesAreOutsideLoop,
1312 std::vector<BasedUser> &UsersToProcess) {
1313 UsersToProcess.reserve(Uses.Users.size());
1314 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1315 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1317 // Move any loop variant operands from the offset field to the immediate
1318 // field of the use, so that we don't try to use something before it is
1320 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1321 UsersToProcess.back().Imm, L, SE);
1322 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1323 "Base value is not loop invariant!");
1326 // We now have a whole bunch of uses of like-strided induction variables, but
1327 // they might all have different bases. We want to emit one PHI node for this
1328 // stride which we fold as many common expressions (between the IVs) into as
1329 // possible. Start by identifying the common expressions in the base values
1330 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1331 // "A+B"), emit it to the preheader, then remove the expression from the
1332 // UsersToProcess base values.
1333 SCEVHandle CommonExprs =
1334 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1336 // Next, figure out what we can represent in the immediate fields of
1337 // instructions. If we can represent anything there, move it to the imm
1338 // fields of the BasedUsers. We do this so that it increases the commonality
1339 // of the remaining uses.
1340 unsigned NumPHI = 0;
1341 bool HasAddress = false;
1342 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1343 // If the user is not in the current loop, this means it is using the exit
1344 // value of the IV. Do not put anything in the base, make sure it's all in
1345 // the immediate field to allow as much factoring as possible.
1346 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1347 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1348 UsersToProcess[i].Base);
1349 UsersToProcess[i].Base =
1350 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1352 // Not all uses are outside the loop.
1353 AllUsesAreOutsideLoop = false;
1355 // Addressing modes can be folded into loads and stores. Be careful that
1356 // the store is through the expression, not of the expression though.
1358 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1359 UsersToProcess[i].OperandValToReplace);
1360 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1368 // If this use isn't an address, then not all uses are addresses.
1369 if (!isAddress && !isPHI)
1370 AllUsesAreAddresses = false;
1372 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1373 UsersToProcess[i].Imm, isAddress, L, SE);
1377 // If one of the use is a PHI node and all other uses are addresses, still
1378 // allow iv reuse. Essentially we are trading one constant multiplication
1379 // for one fewer iv.
1381 AllUsesAreAddresses = false;
1383 // There are no in-loop address uses.
1384 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1385 AllUsesAreAddresses = false;
1390 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1391 /// is valid and profitable for the given set of users of a stride. In
1392 /// full strength-reduction mode, all addresses at the current stride are
1393 /// strength-reduced all the way down to pointer arithmetic.
1395 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1396 const std::vector<BasedUser> &UsersToProcess,
1398 bool AllUsesAreAddresses,
1399 SCEVHandle Stride) {
1400 if (!EnableFullLSRMode)
1403 // The heuristics below aim to avoid increasing register pressure, but
1404 // fully strength-reducing all the addresses increases the number of
1405 // add instructions, so don't do this when optimizing for size.
1406 // TODO: If the loop is large, the savings due to simpler addresses
1407 // may oughtweight the costs of the extra increment instructions.
1408 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1411 // TODO: For now, don't do full strength reduction if there could
1412 // potentially be greater-stride multiples of the current stride
1413 // which could reuse the current stride IV.
1414 if (StrideOrder.back() != Stride)
1417 // Iterate through the uses to find conditions that automatically rule out
1419 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1420 SCEV *Base = UsersToProcess[i].Base;
1421 SCEV *Imm = UsersToProcess[i].Imm;
1422 // If any users have a loop-variant component, they can't be fully
1423 // strength-reduced.
1424 if (Imm && !Imm->isLoopInvariant(L))
1426 // If there are to users with the same base and the difference between
1427 // the two Imm values can't be folded into the address, full
1428 // strength reduction would increase register pressure.
1430 SCEV *CurImm = UsersToProcess[i].Imm;
1431 if ((CurImm || Imm) && CurImm != Imm) {
1432 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1433 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1434 const Instruction *Inst = UsersToProcess[i].Inst;
1435 const Type *UseTy = getAccessType(Inst);
1436 SCEVHandle Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1437 if (!Diff->isZero() &&
1438 (!AllUsesAreAddresses ||
1439 !fitsInAddressMode(Diff, UseTy, TLI, /*HasBaseReg=*/true)))
1442 } while (++i != e && Base == UsersToProcess[i].Base);
1445 // If there's exactly one user in this stride, fully strength-reducing it
1446 // won't increase register pressure. If it's starting from a non-zero base,
1447 // it'll be simpler this way.
1448 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1451 // Otherwise, if there are any users in this stride that don't require
1452 // a register for their base, full strength-reduction will increase
1453 // register pressure.
1454 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1455 if (UsersToProcess[i].Base->isZero())
1458 // Otherwise, go for it.
1462 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1463 /// with the specified start and step values in the specified loop.
1465 /// If NegateStride is true, the stride should be negated by using a
1466 /// subtract instead of an add.
1468 /// Return the created phi node.
1470 static PHINode *InsertAffinePhi(SCEVHandle Start, SCEVHandle Step,
1472 const TargetData *TD,
1473 SCEVExpander &Rewriter) {
1474 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1475 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1477 BasicBlock *Header = L->getHeader();
1478 BasicBlock *Preheader = L->getLoopPreheader();
1479 BasicBlock *LatchBlock = L->getLoopLatch();
1480 const Type *Ty = Start->getType();
1481 if (isa<PointerType>(Ty)) Ty = TD->getIntPtrType();
1483 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1484 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1487 // If the stride is negative, insert a sub instead of an add for the
1489 bool isNegative = isNonConstantNegative(Step);
1490 SCEVHandle IncAmount = Step;
1492 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1494 // Insert an add instruction right before the terminator corresponding
1495 // to the back-edge.
1496 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1497 Preheader->getTerminator());
1500 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1501 LatchBlock->getTerminator());
1503 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1504 LatchBlock->getTerminator());
1506 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1508 PN->addIncoming(IncV, LatchBlock);
1514 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1515 // We want to emit code for users inside the loop first. To do this, we
1516 // rearrange BasedUser so that the entries at the end have
1517 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1518 // vector (so we handle them first).
1519 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1520 PartitionByIsUseOfPostIncrementedValue);
1522 // Sort this by base, so that things with the same base are handled
1523 // together. By partitioning first and stable-sorting later, we are
1524 // guaranteed that within each base we will pop off users from within the
1525 // loop before users outside of the loop with a particular base.
1527 // We would like to use stable_sort here, but we can't. The problem is that
1528 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1529 // we don't have anything to do a '<' comparison on. Because we think the
1530 // number of uses is small, do a horrible bubble sort which just relies on
1532 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1533 // Get a base value.
1534 SCEVHandle Base = UsersToProcess[i].Base;
1536 // Compact everything with this base to be consecutive with this one.
1537 for (unsigned j = i+1; j != e; ++j) {
1538 if (UsersToProcess[j].Base == Base) {
1539 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1546 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1547 /// UsersToProcess, meaning lowering addresses all the way down to direct
1548 /// pointer arithmetic.
1551 LoopStrengthReduce::PrepareToStrengthReduceFully(
1552 std::vector<BasedUser> &UsersToProcess,
1554 SCEVHandle CommonExprs,
1556 SCEVExpander &PreheaderRewriter) {
1557 DOUT << " Fully reducing all users\n";
1559 // Rewrite the UsersToProcess records, creating a separate PHI for each
1560 // unique Base value.
1561 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1562 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1563 // pick the first Imm value here to start with, and adjust it for the
1565 SCEVHandle Imm = UsersToProcess[i].Imm;
1566 SCEVHandle Base = UsersToProcess[i].Base;
1567 SCEVHandle Start = SE->getAddExpr(CommonExprs, Base, Imm);
1568 PHINode *Phi = InsertAffinePhi(Start, Stride, L, TD,
1570 // Loop over all the users with the same base.
1572 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1573 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1574 UsersToProcess[i].Phi = Phi;
1575 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1576 "ShouldUseFullStrengthReductionMode should reject this!");
1577 } while (++i != e && Base == UsersToProcess[i].Base);
1581 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1582 /// given users to share.
1585 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1586 std::vector<BasedUser> &UsersToProcess,
1588 SCEVHandle CommonExprs,
1591 SCEVExpander &PreheaderRewriter) {
1592 DOUT << " Inserting new PHI:\n";
1594 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1598 // Remember this in case a later stride is multiple of this.
1599 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1601 // All the users will share this new IV.
1602 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1603 UsersToProcess[i].Phi = Phi;
1606 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1610 /// PrepareToStrengthReduceWithNewPhi - Prepare for the given users to reuse
1611 /// an induction variable with a stride that is a factor of the current
1612 /// induction variable.
1615 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1616 std::vector<BasedUser> &UsersToProcess,
1618 const IVExpr &ReuseIV,
1619 Instruction *PreInsertPt) {
1620 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1621 << " and BASE " << *ReuseIV.Base << "\n";
1623 // All the users will share the reused IV.
1624 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1625 UsersToProcess[i].Phi = ReuseIV.PHI;
1627 Constant *C = dyn_cast<Constant>(CommonBaseV);
1629 (!C->isNullValue() &&
1630 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1632 // We want the common base emitted into the preheader! This is just
1633 // using cast as a copy so BitCast (no-op cast) is appropriate
1634 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1635 "commonbase", PreInsertPt);
1638 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1639 const Type *AccessTy,
1640 std::vector<BasedUser> &UsersToProcess,
1641 const TargetLowering *TLI) {
1642 SmallVector<Instruction*, 16> AddrModeInsts;
1643 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1644 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1646 ExtAddrMode AddrMode =
1647 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1648 AccessTy, UsersToProcess[i].Inst,
1649 AddrModeInsts, *TLI);
1650 if (GV && GV != AddrMode.BaseGV)
1652 if (Offset && !AddrMode.BaseOffs)
1653 // FIXME: How to accurate check it's immediate offset is folded.
1655 AddrModeInsts.clear();
1660 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1661 /// stride of IV. All of the users may have different starting values, and this
1662 /// may not be the only stride.
1663 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1664 IVUsersOfOneStride &Uses,
1666 // If all the users are moved to another stride, then there is nothing to do.
1667 if (Uses.Users.empty())
1670 // Keep track if every use in UsersToProcess is an address. If they all are,
1671 // we may be able to rewrite the entire collection of them in terms of a
1672 // smaller-stride IV.
1673 bool AllUsesAreAddresses = true;
1675 // Keep track if every use of a single stride is outside the loop. If so,
1676 // we want to be more aggressive about reusing a smaller-stride IV; a
1677 // multiply outside the loop is better than another IV inside. Well, usually.
1678 bool AllUsesAreOutsideLoop = true;
1680 // Transform our list of users and offsets to a bit more complex table. In
1681 // this new vector, each 'BasedUser' contains 'Base' the base of the
1682 // strided accessas well as the old information from Uses. We progressively
1683 // move information from the Base field to the Imm field, until we eventually
1684 // have the full access expression to rewrite the use.
1685 std::vector<BasedUser> UsersToProcess;
1686 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1687 AllUsesAreOutsideLoop,
1690 // Sort the UsersToProcess array so that users with common bases are
1691 // next to each other.
1692 SortUsersToProcess(UsersToProcess);
1694 // If we managed to find some expressions in common, we'll need to carry
1695 // their value in a register and add it in for each use. This will take up
1696 // a register operand, which potentially restricts what stride values are
1698 bool HaveCommonExprs = !CommonExprs->isZero();
1700 const Type *ReplacedTy = CommonExprs->getType();
1701 if (isa<PointerType>(ReplacedTy)) ReplacedTy = TD->getIntPtrType();
1703 // If all uses are addresses, consider sinking the immediate part of the
1704 // common expression back into uses if they can fit in the immediate fields.
1705 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1706 SCEVHandle NewCommon = CommonExprs;
1707 SCEVHandle Imm = SE->getIntegerSCEV(0, ReplacedTy);
1708 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE);
1709 if (!Imm->isZero()) {
1712 // If the immediate part of the common expression is a GV, check if it's
1713 // possible to fold it into the target addressing mode.
1714 GlobalValue *GV = 0;
1715 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1716 GV = dyn_cast<GlobalValue>(SU->getValue());
1718 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1719 Offset = SC->getValue()->getSExtValue();
1721 // Pass VoidTy as the AccessTy to be conservative, because
1722 // there could be multiple access types among all the uses.
1723 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy,
1724 UsersToProcess, TLI);
1727 DOUT << " Sinking " << *Imm << " back down into uses\n";
1728 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1729 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1730 CommonExprs = NewCommon;
1731 HaveCommonExprs = !CommonExprs->isZero();
1737 // Now that we know what we need to do, insert the PHI node itself.
1739 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1741 << " Common base: " << *CommonExprs << "\n";
1743 SCEVExpander Rewriter(*SE, *LI, *TD);
1744 SCEVExpander PreheaderRewriter(*SE, *LI, *TD);
1746 BasicBlock *Preheader = L->getLoopPreheader();
1747 Instruction *PreInsertPt = Preheader->getTerminator();
1748 BasicBlock *LatchBlock = L->getLoopLatch();
1750 Value *CommonBaseV = ConstantInt::get(ReplacedTy, 0);
1752 SCEVHandle RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1753 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1754 SE->getIntegerSCEV(0, Type::Int32Ty),
1757 /// Choose a strength-reduction strategy and prepare for it by creating
1758 /// the necessary PHIs and adjusting the bookkeeping.
1759 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1760 AllUsesAreAddresses, Stride)) {
1761 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1764 // Emit the initial base value into the loop preheader.
1765 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1768 // If all uses are addresses, check if it is possible to reuse an IV with a
1769 // stride that is a factor of this stride. And that the multiple is a number
1770 // that can be encoded in the scale field of the target addressing mode. And
1771 // that we will have a valid instruction after this substition, including
1772 // the immediate field, if any.
1773 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1774 AllUsesAreOutsideLoop,
1775 Stride, ReuseIV, ReplacedTy,
1777 if (isa<SCEVConstant>(RewriteFactor) &&
1778 cast<SCEVConstant>(RewriteFactor)->isZero())
1779 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1780 CommonBaseV, L, PreheaderRewriter);
1782 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1783 ReuseIV, PreInsertPt);
1786 // Process all the users now, replacing their strided uses with
1787 // strength-reduced forms. This outer loop handles all bases, the inner
1788 // loop handles all users of a particular base.
1789 while (!UsersToProcess.empty()) {
1790 SCEVHandle Base = UsersToProcess.back().Base;
1791 Instruction *Inst = UsersToProcess.back().Inst;
1793 // Emit the code for Base into the preheader.
1795 if (!Base->isZero()) {
1796 BaseV = PreheaderRewriter.expandCodeFor(Base, Base->getType(),
1799 DOUT << " INSERTING code for BASE = " << *Base << ":";
1800 if (BaseV->hasName())
1801 DOUT << " Result value name = %" << BaseV->getNameStr();
1804 // If BaseV is a non-zero constant, make sure that it gets inserted into
1805 // the preheader, instead of being forward substituted into the uses. We
1806 // do this by forcing a BitCast (noop cast) to be inserted into the
1807 // preheader in this case.
1808 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false)) {
1809 // We want this constant emitted into the preheader! This is just
1810 // using cast as a copy so BitCast (no-op cast) is appropriate
1811 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1816 // Emit the code to add the immediate offset to the Phi value, just before
1817 // the instructions that we identified as using this stride and base.
1819 // FIXME: Use emitted users to emit other users.
1820 BasedUser &User = UsersToProcess.back();
1822 DOUT << " Examining use ";
1823 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1824 /*PrintType=*/false));
1825 DOUT << " in Inst: " << *Inst;
1827 // If this instruction wants to use the post-incremented value, move it
1828 // after the post-inc and use its value instead of the PHI.
1829 Value *RewriteOp = User.Phi;
1830 if (User.isUseOfPostIncrementedValue) {
1831 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1833 // If this user is in the loop, make sure it is the last thing in the
1834 // loop to ensure it is dominated by the increment.
1835 if (L->contains(User.Inst->getParent()))
1836 User.Inst->moveBefore(LatchBlock->getTerminator());
1839 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1841 if (TD->getTypeSizeInBits(RewriteOp->getType()) !=
1842 TD->getTypeSizeInBits(ReplacedTy)) {
1843 assert(TD->getTypeSizeInBits(RewriteOp->getType()) >
1844 TD->getTypeSizeInBits(ReplacedTy) &&
1845 "Unexpected widening cast!");
1846 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1849 // If we had to insert new instructions for RewriteOp, we have to
1850 // consider that they may not have been able to end up immediately
1851 // next to RewriteOp, because non-PHI instructions may never precede
1852 // PHI instructions in a block. In this case, remember where the last
1853 // instruction was inserted so that if we're replacing a different
1854 // PHI node, we can use the later point to expand the final
1856 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1857 if (RewriteOp == User.Phi) NewBasePt = 0;
1859 // Clear the SCEVExpander's expression map so that we are guaranteed
1860 // to have the code emitted where we expect it.
1863 // If we are reusing the iv, then it must be multiplied by a constant
1864 // factor to take advantage of the addressing mode scale component.
1865 if (!RewriteFactor->isZero()) {
1866 // If we're reusing an IV with a nonzero base (currently this happens
1867 // only when all reuses are outside the loop) subtract that base here.
1868 // The base has been used to initialize the PHI node but we don't want
1870 if (!ReuseIV.Base->isZero()) {
1871 SCEVHandle typedBase = ReuseIV.Base;
1872 if (RewriteExpr->getType()->getPrimitiveSizeInBits() !=
1873 ReuseIV.Base->getType()->getPrimitiveSizeInBits()) {
1874 // It's possible the original IV is a larger type than the new IV,
1875 // in which case we have to truncate the Base. We checked in
1876 // RequiresTypeConversion that this is valid.
1877 assert (RewriteExpr->getType()->getPrimitiveSizeInBits() <
1878 ReuseIV.Base->getType()->getPrimitiveSizeInBits() &&
1879 "Unexpected lengthening conversion!");
1880 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1881 RewriteExpr->getType());
1883 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1886 // Multiply old variable, with base removed, by new scale factor.
1887 RewriteExpr = SE->getMulExpr(RewriteFactor,
1890 // The common base is emitted in the loop preheader. But since we
1891 // are reusing an IV, it has not been used to initialize the PHI node.
1892 // Add it to the expression used to rewrite the uses.
1893 // When this use is outside the loop, we earlier subtracted the
1894 // common base, and are adding it back here. Use the same expression
1895 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1896 if (!CommonExprs->isZero()) {
1897 if (L->contains(User.Inst->getParent()))
1898 RewriteExpr = SE->getAddExpr(RewriteExpr,
1899 SE->getUnknown(CommonBaseV));
1901 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1905 // Now that we know what we need to do, insert code before User for the
1906 // immediate and any loop-variant expressions.
1908 // Add BaseV to the PHI value if needed.
1909 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1911 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1915 // Mark old value we replaced as possibly dead, so that it is eliminated
1916 // if we just replaced the last use of that value.
1917 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1919 UsersToProcess.pop_back();
1922 // If there are any more users to process with the same base, process them
1923 // now. We sorted by base above, so we just have to check the last elt.
1924 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1925 // TODO: Next, find out which base index is the most common, pull it out.
1928 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1929 // different starting values, into different PHIs.
1932 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1933 /// set the IV user and stride information and return true, otherwise return
1935 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1936 const SCEVHandle *&CondStride) {
1937 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1939 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1940 IVUsesByStride.find(StrideOrder[Stride]);
1941 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1943 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1944 E = SI->second.Users.end(); UI != E; ++UI)
1945 if (UI->User == Cond) {
1946 // NOTE: we could handle setcc instructions with multiple uses here, but
1947 // InstCombine does it as well for simple uses, it's not clear that it
1948 // occurs enough in real life to handle.
1950 CondStride = &SI->first;
1958 // Constant strides come first which in turns are sorted by their absolute
1959 // values. If absolute values are the same, then positive strides comes first.
1961 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1962 struct StrideCompare {
1963 const TargetData *TD;
1964 explicit StrideCompare(const TargetData *td) : TD(td) {}
1966 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1967 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1968 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1970 int64_t LV = LHSC->getValue()->getSExtValue();
1971 int64_t RV = RHSC->getValue()->getSExtValue();
1972 uint64_t ALV = (LV < 0) ? -LV : LV;
1973 uint64_t ARV = (RV < 0) ? -RV : RV;
1981 // If it's the same value but different type, sort by bit width so
1982 // that we emit larger induction variables before smaller
1983 // ones, letting the smaller be re-written in terms of larger ones.
1984 return TD->getTypeSizeInBits(RHS->getType()) <
1985 TD->getTypeSizeInBits(LHS->getType());
1987 return LHSC && !RHSC;
1992 /// ChangeCompareStride - If a loop termination compare instruction is the
1993 /// only use of its stride, and the compaison is against a constant value,
1994 /// try eliminate the stride by moving the compare instruction to another
1995 /// stride and change its constant operand accordingly. e.g.
2001 /// if (v2 < 10) goto loop
2006 /// if (v1 < 30) goto loop
2007 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
2008 IVStrideUse* &CondUse,
2009 const SCEVHandle* &CondStride) {
2010 if (StrideOrder.size() < 2 ||
2011 IVUsesByStride[*CondStride].Users.size() != 1)
2013 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
2014 if (!SC) return Cond;
2016 ICmpInst::Predicate Predicate = Cond->getPredicate();
2017 int64_t CmpSSInt = SC->getValue()->getSExtValue();
2018 unsigned BitWidth = TD->getTypeSizeInBits((*CondStride)->getType());
2019 uint64_t SignBit = 1ULL << (BitWidth-1);
2020 const Type *CmpTy = Cond->getOperand(0)->getType();
2021 const Type *NewCmpTy = NULL;
2022 unsigned TyBits = TD->getTypeSizeInBits(CmpTy);
2023 unsigned NewTyBits = 0;
2024 SCEVHandle *NewStride = NULL;
2025 Value *NewCmpLHS = NULL;
2026 Value *NewCmpRHS = NULL;
2028 SCEVHandle NewOffset = SE->getIntegerSCEV(0, UIntPtrTy);
2030 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
2031 int64_t CmpVal = C->getValue().getSExtValue();
2033 // Check stride constant and the comparision constant signs to detect
2035 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
2038 // Look for a suitable stride / iv as replacement.
2039 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
2040 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2041 IVUsesByStride.find(StrideOrder[i]);
2042 if (!isa<SCEVConstant>(SI->first))
2044 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2045 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
2048 Scale = SSInt / CmpSSInt;
2049 int64_t NewCmpVal = CmpVal * Scale;
2050 APInt Mul = APInt(BitWidth, NewCmpVal);
2051 // Check for overflow.
2052 if (Mul.getSExtValue() != NewCmpVal)
2055 // Watch out for overflow.
2056 if (ICmpInst::isSignedPredicate(Predicate) &&
2057 (CmpVal & SignBit) != (NewCmpVal & SignBit))
2060 if (NewCmpVal == CmpVal)
2062 // Pick the best iv to use trying to avoid a cast.
2064 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2065 E = SI->second.Users.end(); UI != E; ++UI) {
2066 NewCmpLHS = UI->OperandValToReplace;
2067 if (NewCmpLHS->getType() == CmpTy)
2073 NewCmpTy = NewCmpLHS->getType();
2074 NewTyBits = TD->getTypeSizeInBits(NewCmpTy);
2075 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2076 // Check if it is possible to rewrite it using
2077 // an iv / stride of a smaller integer type.
2078 bool TruncOk = false;
2079 if (NewCmpTy->isInteger()) {
2080 unsigned Bits = NewTyBits;
2081 if (ICmpInst::isSignedPredicate(Predicate))
2083 uint64_t Mask = (1ULL << Bits) - 1;
2084 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
2091 // Don't rewrite if use offset is non-constant and the new type is
2092 // of a different type.
2093 // FIXME: too conservative?
2094 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset))
2097 bool AllUsesAreAddresses = true;
2098 bool AllUsesAreOutsideLoop = true;
2099 std::vector<BasedUser> UsersToProcess;
2100 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
2101 AllUsesAreAddresses,
2102 AllUsesAreOutsideLoop,
2104 // Avoid rewriting the compare instruction with an iv of new stride
2105 // if it's likely the new stride uses will be rewritten using the
2106 // stride of the compare instruction.
2107 if (AllUsesAreAddresses &&
2108 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess))
2111 // If scale is negative, use swapped predicate unless it's testing
2113 if (Scale < 0 && !Cond->isEquality())
2114 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2116 NewStride = &StrideOrder[i];
2117 if (!isa<PointerType>(NewCmpTy))
2118 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2120 ConstantInt *CI = ConstantInt::get(UIntPtrTy, NewCmpVal);
2121 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2123 NewOffset = TyBits == NewTyBits
2124 ? SE->getMulExpr(CondUse->Offset,
2125 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
2126 : SE->getConstant(ConstantInt::get(NewCmpTy,
2127 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
2132 // Forgo this transformation if it the increment happens to be
2133 // unfortunately positioned after the condition, and the condition
2134 // has multiple uses which prevent it from being moved immediately
2135 // before the branch. See
2136 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2137 // for an example of this situation.
2138 if (!Cond->hasOneUse()) {
2139 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2146 // Create a new compare instruction using new stride / iv.
2147 ICmpInst *OldCond = Cond;
2148 // Insert new compare instruction.
2149 Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
2150 L->getHeader()->getName() + ".termcond",
2153 // Remove the old compare instruction. The old indvar is probably dead too.
2154 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
2155 SE->deleteValueFromRecords(OldCond);
2156 OldCond->replaceAllUsesWith(Cond);
2157 OldCond->eraseFromParent();
2159 IVUsesByStride[*CondStride].Users.pop_back();
2160 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewCmpLHS);
2161 CondUse = &IVUsesByStride[*NewStride].Users.back();
2162 CondStride = NewStride;
2169 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
2170 /// an smax computation.
2172 /// This is a narrow solution to a specific, but acute, problem. For loops
2178 /// } while (++i < n);
2180 /// where the comparison is signed, the trip count isn't just 'n', because
2181 /// 'n' could be negative. And unfortunately this can come up even for loops
2182 /// where the user didn't use a C do-while loop. For example, seemingly
2183 /// well-behaved top-test loops will commonly be lowered like this:
2189 /// } while (++i < n);
2192 /// and then it's possible for subsequent optimization to obscure the if
2193 /// test in such a way that indvars can't find it.
2195 /// When indvars can't find the if test in loops like this, it creates a
2196 /// signed-max expression, which allows it to give the loop a canonical
2197 /// induction variable:
2200 /// smax = n < 1 ? 1 : n;
2203 /// } while (++i != smax);
2205 /// Canonical induction variables are necessary because the loop passes
2206 /// are designed around them. The most obvious example of this is the
2207 /// LoopInfo analysis, which doesn't remember trip count values. It
2208 /// expects to be able to rediscover the trip count each time it is
2209 /// needed, and it does this using a simple analyis that only succeeds if
2210 /// the loop has a canonical induction variable.
2212 /// However, when it comes time to generate code, the maximum operation
2213 /// can be quite costly, especially if it's inside of an outer loop.
2215 /// This function solves this problem by detecting this type of loop and
2216 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2217 /// the instructions for the maximum computation.
2219 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
2220 IVStrideUse* &CondUse) {
2221 // Check that the loop matches the pattern we're looking for.
2222 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2223 Cond->getPredicate() != CmpInst::ICMP_NE)
2226 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2227 if (!Sel || !Sel->hasOneUse()) return Cond;
2229 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2230 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2232 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2234 // Add one to the backedge-taken count to get the trip count.
2235 SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2237 // Check for a max calculation that matches the pattern.
2238 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
2239 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
2241 SCEVHandle SMaxLHS = SMax->getOperand(0);
2242 SCEVHandle SMaxRHS = SMax->getOperand(1);
2243 if (!SMaxLHS || SMaxLHS != One) return Cond;
2245 // Check the relevant induction variable for conformance to
2247 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2248 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2249 if (!AR || !AR->isAffine() ||
2250 AR->getStart() != One ||
2251 AR->getStepRecurrence(*SE) != One)
2254 assert(AR->getLoop() == L &&
2255 "Loop condition operand is an addrec in a different loop!");
2257 // Check the right operand of the select, and remember it, as it will
2258 // be used in the new comparison instruction.
2260 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
2261 NewRHS = Sel->getOperand(1);
2262 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
2263 NewRHS = Sel->getOperand(2);
2264 if (!NewRHS) return Cond;
2266 // Ok, everything looks ok to change the condition into an SLT or SGE and
2267 // delete the max calculation.
2269 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2272 Cond->getOperand(0), NewRHS, "scmp", Cond);
2274 // Delete the max calculation instructions.
2275 SE->deleteValueFromRecords(Cond);
2276 Cond->replaceAllUsesWith(NewCond);
2277 Cond->eraseFromParent();
2278 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2279 SE->deleteValueFromRecords(Sel);
2280 Sel->eraseFromParent();
2281 if (Cmp->use_empty()) {
2282 SE->deleteValueFromRecords(Cmp);
2283 Cmp->eraseFromParent();
2285 CondUse->User = NewCond;
2289 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2290 /// inside the loop then try to eliminate the cast opeation.
2291 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2293 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2294 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2297 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2299 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2300 IVUsesByStride.find(StrideOrder[Stride]);
2301 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2302 if (!isa<SCEVConstant>(SI->first))
2305 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2306 E = SI->second.Users.end(); UI != E; /* empty */) {
2307 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2309 Instruction *ShadowUse = CandidateUI->User;
2310 const Type *DestTy = NULL;
2312 /* If shadow use is a int->float cast then insert a second IV
2313 to eliminate this cast.
2315 for (unsigned i = 0; i < n; ++i)
2321 for (unsigned i = 0; i < n; ++i, ++d)
2324 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2325 DestTy = UCast->getDestTy();
2326 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2327 DestTy = SCast->getDestTy();
2328 if (!DestTy) continue;
2331 /* If target does not support DestTy natively then do not apply
2332 this transformation. */
2333 MVT DVT = TLI->getValueType(DestTy);
2334 if (!TLI->isTypeLegal(DVT)) continue;
2337 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2339 if (PH->getNumIncomingValues() != 2) continue;
2341 const Type *SrcTy = PH->getType();
2342 int Mantissa = DestTy->getFPMantissaWidth();
2343 if (Mantissa == -1) continue;
2344 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
2347 unsigned Entry, Latch;
2348 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2356 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2357 if (!Init) continue;
2358 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2360 BinaryOperator *Incr =
2361 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2362 if (!Incr) continue;
2363 if (Incr->getOpcode() != Instruction::Add
2364 && Incr->getOpcode() != Instruction::Sub)
2367 /* Initialize new IV, double d = 0.0 in above example. */
2368 ConstantInt *C = NULL;
2369 if (Incr->getOperand(0) == PH)
2370 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2371 else if (Incr->getOperand(1) == PH)
2372 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2378 /* Add new PHINode. */
2379 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2381 /* create new increment. '++d' in above example. */
2382 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2383 BinaryOperator *NewIncr =
2384 BinaryOperator::Create(Incr->getOpcode(),
2385 NewPH, CFP, "IV.S.next.", Incr);
2387 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2388 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2390 /* Remove cast operation */
2391 SE->deleteValueFromRecords(ShadowUse);
2392 ShadowUse->replaceAllUsesWith(NewPH);
2393 ShadowUse->eraseFromParent();
2394 SI->second.Users.erase(CandidateUI);
2401 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2402 // uses in the loop, look to see if we can eliminate some, in favor of using
2403 // common indvars for the different uses.
2404 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2405 // TODO: implement optzns here.
2407 OptimizeShadowIV(L);
2409 // Finally, get the terminating condition for the loop if possible. If we
2410 // can, we want to change it to use a post-incremented version of its
2411 // induction variable, to allow coalescing the live ranges for the IV into
2412 // one register value.
2413 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2414 BasicBlock *Preheader = L->getLoopPreheader();
2415 BasicBlock *LatchBlock =
2416 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2417 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2418 if (!TermBr || TermBr->isUnconditional() ||
2419 !isa<ICmpInst>(TermBr->getCondition()))
2421 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2423 // Search IVUsesByStride to find Cond's IVUse if there is one.
2424 IVStrideUse *CondUse = 0;
2425 const SCEVHandle *CondStride = 0;
2427 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2428 return; // setcc doesn't use the IV.
2430 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2431 // being unable to find a sufficient guard, for example), change the loop
2432 // comparison to use SLT instead of NE.
2433 Cond = OptimizeSMax(L, Cond, CondUse);
2435 // If possible, change stride and operands of the compare instruction to
2436 // eliminate one stride.
2437 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2439 // It's possible for the setcc instruction to be anywhere in the loop, and
2440 // possible for it to have multiple users. If it is not immediately before
2441 // the latch block branch, move it.
2442 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2443 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2444 Cond->moveBefore(TermBr);
2446 // Otherwise, clone the terminating condition and insert into the loopend.
2447 Cond = cast<ICmpInst>(Cond->clone());
2448 Cond->setName(L->getHeader()->getName() + ".termcond");
2449 LatchBlock->getInstList().insert(TermBr, Cond);
2451 // Clone the IVUse, as the old use still exists!
2452 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2453 CondUse->OperandValToReplace);
2454 CondUse = &IVUsesByStride[*CondStride].Users.back();
2458 // If we get to here, we know that we can transform the setcc instruction to
2459 // use the post-incremented version of the IV, allowing us to coalesce the
2460 // live ranges for the IV correctly.
2461 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2462 CondUse->isUseOfPostIncrementedValue = true;
2466 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2468 LI = &getAnalysis<LoopInfo>();
2469 DT = &getAnalysis<DominatorTree>();
2470 SE = &getAnalysis<ScalarEvolution>();
2471 TD = &getAnalysis<TargetData>();
2472 UIntPtrTy = TD->getIntPtrType();
2475 // Find all uses of induction variables in this loop, and categorize
2476 // them by stride. Start by finding all of the PHI nodes in the header for
2477 // this loop. If they are induction variables, inspect their uses.
2478 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2479 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2480 AddUsersIfInteresting(I, L, Processed);
2482 if (!IVUsesByStride.empty()) {
2484 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2489 // Sort the StrideOrder so we process larger strides first.
2490 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare(TD));
2492 // Optimize induction variables. Some indvar uses can be transformed to use
2493 // strides that will be needed for other purposes. A common example of this
2494 // is the exit test for the loop, which can often be rewritten to use the
2495 // computation of some other indvar to decide when to terminate the loop.
2498 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2499 // doing computation in byte values, promote to 32-bit values if safe.
2501 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2502 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2503 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2504 // Need to be careful that IV's are all the same type. Only works for
2505 // intptr_t indvars.
2507 // IVsByStride keeps IVs for one particular loop.
2508 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2510 // Note: this processes each stride/type pair individually. All users
2511 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2512 // Also, note that we iterate over IVUsesByStride indirectly by using
2513 // StrideOrder. This extra layer of indirection makes the ordering of
2514 // strides deterministic - not dependent on map order.
2515 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2516 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2517 IVUsesByStride.find(StrideOrder[Stride]);
2518 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2519 StrengthReduceStridedIVUsers(SI->first, SI->second, L);
2523 // We're done analyzing this loop; release all the state we built up for it.
2524 IVUsesByStride.clear();
2525 IVsByStride.clear();
2526 StrideOrder.clear();
2528 // Clean up after ourselves
2529 if (!DeadInsts.empty()) {
2530 DeleteTriviallyDeadInstructions();
2532 BasicBlock::iterator I = L->getHeader()->begin();
2533 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2534 // At this point, we know that we have killed one or more IV users.
2535 // It is worth checking to see if the cannonical indvar is also
2536 // dead, so that we can remove it as well.
2538 // We can remove a PHI if it is on a cycle in the def-use graph
2539 // where each node in the cycle has degree one, i.e. only one use,
2540 // and is an instruction with no side effects.
2542 // FIXME: this needs to eliminate an induction variable even if it's being
2543 // compared against some value to decide loop termination.
2544 if (!PN->hasOneUse())
2547 SmallPtrSet<PHINode *, 4> PHIs;
2548 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2549 J && J->hasOneUse() && !J->mayWriteToMemory();
2550 J = dyn_cast<Instruction>(*J->use_begin())) {
2551 // If we find the original PHI, we've discovered a cycle.
2553 // Break the cycle and mark the PHI for deletion.
2554 SE->deleteValueFromRecords(PN);
2555 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2556 DeadInsts.push_back(PN);
2560 // If we find a PHI more than once, we're on a cycle that
2561 // won't prove fruitful.
2562 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2566 DeleteTriviallyDeadInstructions();