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/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Support/CFG.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/ValueHandle.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 {
117 /// IVUsesByStride - Keep track of all uses of induction variables that we
118 /// are interested in. The key of the map is the stride of the access.
119 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
121 /// IVsByStride - Keep track of all IVs that have been inserted for a
122 /// particular stride.
123 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
125 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
126 /// We use this to iterate over the IVUsesByStride collection without being
127 /// dependent on random ordering of pointers in the process.
128 SmallVector<SCEVHandle, 16> StrideOrder;
130 /// DeadInsts - Keep track of instructions we may have made dead, so that
131 /// we can remove them after we are done working.
132 SmallVector<Instruction*, 16> DeadInsts;
134 /// TLI - Keep a pointer of a TargetLowering to consult for determining
135 /// transformation profitability.
136 const TargetLowering *TLI;
139 static char ID; // Pass ID, replacement for typeid
140 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
141 LoopPass(&ID), TLI(tli) {
144 bool runOnLoop(Loop *L, LPPassManager &LPM);
146 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
147 // We split critical edges, so we change the CFG. However, we do update
148 // many analyses if they are around.
149 AU.addPreservedID(LoopSimplifyID);
150 AU.addPreserved<LoopInfo>();
151 AU.addPreserved<DominanceFrontier>();
152 AU.addPreserved<DominatorTree>();
154 AU.addRequiredID(LoopSimplifyID);
155 AU.addRequired<LoopInfo>();
156 AU.addRequired<DominatorTree>();
157 AU.addRequired<ScalarEvolution>();
158 AU.addPreserved<ScalarEvolution>();
162 bool AddUsersIfInteresting(Instruction *I, Loop *L,
163 SmallPtrSet<Instruction*,16> &Processed);
164 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
165 IVStrideUse* &CondUse,
166 const SCEVHandle* &CondStride);
168 void OptimizeIndvars(Loop *L);
169 void OptimizeLoopCountIV(Loop *L);
170 void OptimizeLoopTermCond(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 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
260 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
263 DeadInsts.push_back(U);
267 I->eraseFromParent();
272 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
273 /// subexpression that is an AddRec from a loop other than L. An outer loop
274 /// of L is OK, but not an inner loop nor a disjoint loop.
275 static bool containsAddRecFromDifferentLoop(SCEVHandle S, Loop *L) {
276 // This is very common, put it first.
277 if (isa<SCEVConstant>(S))
279 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
280 for (unsigned int i=0; i< AE->getNumOperands(); i++)
281 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
285 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
286 if (const Loop *newLoop = AE->getLoop()) {
289 // if newLoop is an outer loop of L, this is OK.
290 if (!LoopInfoBase<BasicBlock>::isNotAlreadyContainedIn(L, newLoop))
295 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
296 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
297 containsAddRecFromDifferentLoop(DE->getRHS(), L);
299 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
300 // need this when it is.
301 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
302 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
303 containsAddRecFromDifferentLoop(DE->getRHS(), L);
305 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
306 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
310 /// getSCEVStartAndStride - Compute the start and stride of this expression,
311 /// returning false if the expression is not a start/stride pair, or true if it
312 /// is. The stride must be a loop invariant expression, but the start may be
313 /// a mix of loop invariant and loop variant expressions. The start cannot,
314 /// however, contain an AddRec from a different loop, unless that loop is an
315 /// outer loop of the current loop.
316 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
317 SCEVHandle &Start, SCEVHandle &Stride,
318 ScalarEvolution *SE, DominatorTree *DT) {
319 SCEVHandle TheAddRec = Start; // Initialize to zero.
321 // If the outer level is an AddExpr, the operands are all start values except
322 // for a nested AddRecExpr.
323 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
324 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
325 if (const SCEVAddRecExpr *AddRec =
326 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
327 if (AddRec->getLoop() == L)
328 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
330 return false; // Nested IV of some sort?
332 Start = SE->getAddExpr(Start, AE->getOperand(i));
335 } else if (isa<SCEVAddRecExpr>(SH)) {
338 return false; // not analyzable.
341 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
342 if (!AddRec || AddRec->getLoop() != L) return false;
344 // FIXME: Generalize to non-affine IV's.
345 if (!AddRec->isAffine()) return false;
347 // If Start contains an SCEVAddRecExpr from a different loop, other than an
348 // outer loop of the current loop, reject it. SCEV has no concept of
349 // operating on more than one loop at a time so don't confuse it with such
351 if (containsAddRecFromDifferentLoop(AddRec->getOperand(0), L))
354 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
356 if (!isa<SCEVConstant>(AddRec->getOperand(1))) {
357 // If stride is an instruction, make sure it dominates the loop preheader.
358 // Otherwise we could end up with a use before def situation.
359 BasicBlock *Preheader = L->getLoopPreheader();
360 if (!AddRec->getOperand(1)->dominates(Preheader, DT))
363 DOUT << "[" << L->getHeader()->getName()
364 << "] Variable stride: " << *AddRec << "\n";
367 Stride = AddRec->getOperand(1);
371 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
372 /// and now we need to decide whether the user should use the preinc or post-inc
373 /// value. If this user should use the post-inc version of the IV, return true.
375 /// Choosing wrong here can break dominance properties (if we choose to use the
376 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
377 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
378 /// should use the post-inc value).
379 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
380 Loop *L, DominatorTree *DT, Pass *P,
381 SmallVectorImpl<Instruction*> &DeadInsts){
382 // If the user is in the loop, use the preinc value.
383 if (L->contains(User->getParent())) return false;
385 BasicBlock *LatchBlock = L->getLoopLatch();
387 // Ok, the user is outside of the loop. If it is dominated by the latch
388 // block, use the post-inc value.
389 if (DT->dominates(LatchBlock, User->getParent()))
392 // There is one case we have to be careful of: PHI nodes. These little guys
393 // can live in blocks that do not dominate the latch block, but (since their
394 // uses occur in the predecessor block, not the block the PHI lives in) should
395 // still use the post-inc value. Check for this case now.
396 PHINode *PN = dyn_cast<PHINode>(User);
397 if (!PN) return false; // not a phi, not dominated by latch block.
399 // Look at all of the uses of IV by the PHI node. If any use corresponds to
400 // a block that is not dominated by the latch block, give up and use the
401 // preincremented value.
402 unsigned NumUses = 0;
403 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
404 if (PN->getIncomingValue(i) == IV) {
406 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
410 // Okay, all uses of IV by PN are in predecessor blocks that really are
411 // dominated by the latch block. Use the post-incremented value.
415 /// isAddressUse - Returns true if the specified instruction is using the
416 /// specified value as an address.
417 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
418 bool isAddress = isa<LoadInst>(Inst);
419 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
420 if (SI->getOperand(1) == OperandVal)
422 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
423 // Addressing modes can also be folded into prefetches and a variety
425 switch (II->getIntrinsicID()) {
427 case Intrinsic::prefetch:
428 case Intrinsic::x86_sse2_loadu_dq:
429 case Intrinsic::x86_sse2_loadu_pd:
430 case Intrinsic::x86_sse_loadu_ps:
431 case Intrinsic::x86_sse_storeu_ps:
432 case Intrinsic::x86_sse2_storeu_pd:
433 case Intrinsic::x86_sse2_storeu_dq:
434 case Intrinsic::x86_sse2_storel_dq:
435 if (II->getOperand(1) == OperandVal)
443 /// getAccessType - Return the type of the memory being accessed.
444 static const Type *getAccessType(const Instruction *Inst) {
445 const Type *UseTy = Inst->getType();
446 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
447 UseTy = SI->getOperand(0)->getType();
448 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
449 // Addressing modes can also be folded into prefetches and a variety
451 switch (II->getIntrinsicID()) {
453 case Intrinsic::x86_sse_storeu_ps:
454 case Intrinsic::x86_sse2_storeu_pd:
455 case Intrinsic::x86_sse2_storeu_dq:
456 case Intrinsic::x86_sse2_storel_dq:
457 UseTy = II->getOperand(1)->getType();
464 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
465 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
466 /// return true. Otherwise, return false.
467 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
468 SmallPtrSet<Instruction*,16> &Processed) {
469 if (!SE->isSCEVable(I->getType()))
470 return false; // Void and FP expressions cannot be reduced.
472 // LSR is not APInt clean, do not touch integers bigger than 64-bits.
473 if (SE->getTypeSizeInBits(I->getType()) > 64)
476 if (!Processed.insert(I))
477 return true; // Instruction already handled.
479 // Get the symbolic expression for this instruction.
480 SCEVHandle ISE = SE->getSCEV(I);
481 if (isa<SCEVCouldNotCompute>(ISE)) return false;
483 // Get the start and stride for this expression.
484 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
485 SCEVHandle Stride = Start;
486 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE, DT))
487 return false; // Non-reducible symbolic expression, bail out.
489 std::vector<Instruction *> IUsers;
490 // Collect all I uses now because IVUseShouldUsePostIncValue may
491 // invalidate use_iterator.
492 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
493 IUsers.push_back(cast<Instruction>(*UI));
495 for (unsigned iused_index = 0, iused_size = IUsers.size();
496 iused_index != iused_size; ++iused_index) {
498 Instruction *User = IUsers[iused_index];
500 // Do not infinitely recurse on PHI nodes.
501 if (isa<PHINode>(User) && Processed.count(User))
504 // Descend recursively, but not into PHI nodes outside the current loop.
505 // It's important to see the entire expression outside the loop to get
506 // choices that depend on addressing mode use right, although we won't
507 // consider references ouside the loop in all cases.
508 // If User is already in Processed, we don't want to recurse into it again,
509 // but do want to record a second reference in the same instruction.
510 bool AddUserToIVUsers = false;
511 if (LI->getLoopFor(User->getParent()) != L) {
512 if (isa<PHINode>(User) || Processed.count(User) ||
513 !AddUsersIfInteresting(User, L, Processed)) {
514 DOUT << "FOUND USER in other loop: " << *User
515 << " OF SCEV: " << *ISE << "\n";
516 AddUserToIVUsers = true;
518 } else if (Processed.count(User) ||
519 !AddUsersIfInteresting(User, L, Processed)) {
520 DOUT << "FOUND USER: " << *User
521 << " OF SCEV: " << *ISE << "\n";
522 AddUserToIVUsers = true;
525 if (AddUserToIVUsers) {
526 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
527 if (StrideUses.Users.empty()) // First occurrence of this stride?
528 StrideOrder.push_back(Stride);
530 // Okay, we found a user that we cannot reduce. Analyze the instruction
531 // and decide what to do with it. If we are a use inside of the loop, use
532 // the value before incrementation, otherwise use it after incrementation.
533 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
534 // The value used will be incremented by the stride more than we are
535 // expecting, so subtract this off.
536 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
537 StrideUses.addUser(NewStart, User, I);
538 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
539 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
541 StrideUses.addUser(Start, User, I);
549 /// BasedUser - For a particular base value, keep information about how we've
550 /// partitioned the expression so far.
552 /// SE - The current ScalarEvolution object.
555 /// Base - The Base value for the PHI node that needs to be inserted for
556 /// this use. As the use is processed, information gets moved from this
557 /// field to the Imm field (below). BasedUser values are sorted by this
561 /// Inst - The instruction using the induction variable.
564 /// OperandValToReplace - The operand value of Inst to replace with the
566 Value *OperandValToReplace;
568 /// Imm - The immediate value that should be added to the base immediately
569 /// before Inst, because it will be folded into the imm field of the
570 /// instruction. This is also sometimes used for loop-variant values that
571 /// must be added inside the loop.
574 /// Phi - The induction variable that performs the striding that
575 /// should be used for this user.
578 // isUseOfPostIncrementedValue - True if this should use the
579 // post-incremented version of this IV, not the preincremented version.
580 // This can only be set in special cases, such as the terminating setcc
581 // instruction for a loop and uses outside the loop that are dominated by
583 bool isUseOfPostIncrementedValue;
585 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
586 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
587 OperandValToReplace(IVSU.OperandValToReplace),
588 Imm(SE->getIntegerSCEV(0, Base->getType())),
589 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
591 // Once we rewrite the code to insert the new IVs we want, update the
592 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
594 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
595 Instruction *InsertPt,
596 SCEVExpander &Rewriter, Loop *L, Pass *P,
597 SmallVectorImpl<Instruction*> &DeadInsts);
599 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
601 SCEVExpander &Rewriter,
602 Instruction *IP, Loop *L);
607 void BasedUser::dump() const {
608 cerr << " Base=" << *Base;
609 cerr << " Imm=" << *Imm;
610 cerr << " Inst: " << *Inst;
613 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
615 SCEVExpander &Rewriter,
616 Instruction *IP, Loop *L) {
617 // Figure out where we *really* want to insert this code. In particular, if
618 // the user is inside of a loop that is nested inside of L, we really don't
619 // want to insert this expression before the user, we'd rather pull it out as
620 // many loops as possible.
621 LoopInfo &LI = Rewriter.getLoopInfo();
622 Instruction *BaseInsertPt = IP;
624 // Figure out the most-nested loop that IP is in.
625 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
627 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
628 // the preheader of the outer-most loop where NewBase is not loop invariant.
629 if (L->contains(IP->getParent()))
630 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
631 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
632 InsertLoop = InsertLoop->getParentLoop();
635 Value *Base = Rewriter.expandCodeFor(NewBase, Ty, BaseInsertPt);
637 // If there is no immediate value, skip the next part.
641 // If we are inserting the base and imm values in the same block, make sure to
642 // adjust the IP position if insertion reused a result.
643 if (IP == BaseInsertPt)
644 IP = Rewriter.getInsertionPoint();
646 // Always emit the immediate (if non-zero) into the same block as the user.
647 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
648 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
652 // Once we rewrite the code to insert the new IVs we want, update the
653 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
654 // to it. NewBasePt is the last instruction which contributes to the
655 // value of NewBase in the case that it's a diffferent instruction from
656 // the PHI that NewBase is computed from, or null otherwise.
658 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
659 Instruction *NewBasePt,
660 SCEVExpander &Rewriter, Loop *L, Pass *P,
661 SmallVectorImpl<Instruction*> &DeadInsts){
662 if (!isa<PHINode>(Inst)) {
663 // By default, insert code at the user instruction.
664 BasicBlock::iterator InsertPt = Inst;
666 // However, if the Operand is itself an instruction, the (potentially
667 // complex) inserted code may be shared by many users. Because of this, we
668 // want to emit code for the computation of the operand right before its old
669 // computation. This is usually safe, because we obviously used to use the
670 // computation when it was computed in its current block. However, in some
671 // cases (e.g. use of a post-incremented induction variable) the NewBase
672 // value will be pinned to live somewhere after the original computation.
673 // In this case, we have to back off.
675 // If this is a use outside the loop (which means after, since it is based
676 // on a loop indvar) we use the post-incremented value, so that we don't
677 // artificially make the preinc value live out the bottom of the loop.
678 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
679 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
680 InsertPt = NewBasePt;
682 } else if (Instruction *OpInst
683 = dyn_cast<Instruction>(OperandValToReplace)) {
685 while (isa<PHINode>(InsertPt)) ++InsertPt;
688 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
689 OperandValToReplace->getType(),
690 Rewriter, InsertPt, L);
691 // Replace the use of the operand Value with the new Phi we just created.
692 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
694 DOUT << " Replacing with ";
695 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
696 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
700 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
701 // expression into each operand block that uses it. Note that PHI nodes can
702 // have multiple entries for the same predecessor. We use a map to make sure
703 // that a PHI node only has a single Value* for each predecessor (which also
704 // prevents us from inserting duplicate code in some blocks).
705 DenseMap<BasicBlock*, Value*> InsertedCode;
706 PHINode *PN = cast<PHINode>(Inst);
707 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
708 if (PN->getIncomingValue(i) == OperandValToReplace) {
709 // If the original expression is outside the loop, put the replacement
710 // code in the same place as the original expression,
711 // which need not be an immediate predecessor of this PHI. This way we
712 // need only one copy of it even if it is referenced multiple times in
713 // the PHI. We don't do this when the original expression is inside the
714 // loop because multiple copies sometimes do useful sinking of code in
716 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
717 if (L->contains(OldLoc->getParent())) {
718 // If this is a critical edge, split the edge so that we do not insert
719 // the code on all predecessor/successor paths. We do this unless this
720 // is the canonical backedge for this loop, as this can make some
721 // inserted code be in an illegal position.
722 BasicBlock *PHIPred = PN->getIncomingBlock(i);
723 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
724 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
726 // First step, split the critical edge.
727 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
729 // Next step: move the basic block. In particular, if the PHI node
730 // is outside of the loop, and PredTI is in the loop, we want to
731 // move the block to be immediately before the PHI block, not
732 // immediately after PredTI.
733 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
734 BasicBlock *NewBB = PN->getIncomingBlock(i);
735 NewBB->moveBefore(PN->getParent());
738 // Splitting the edge can reduce the number of PHI entries we have.
739 e = PN->getNumIncomingValues();
742 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
744 // Insert the code into the end of the predecessor block.
745 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
746 PN->getIncomingBlock(i)->getTerminator() :
747 OldLoc->getParent()->getTerminator();
748 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
749 Rewriter, InsertPt, L);
751 DOUT << " Changing PHI use to ";
752 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
753 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
756 // Replace the use of the operand Value with the new Phi we just created.
757 PN->setIncomingValue(i, Code);
762 // PHI node might have become a constant value after SplitCriticalEdge.
763 DeadInsts.push_back(Inst);
767 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
768 /// mode, and does not need to be put in a register first.
769 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
770 const TargetLowering *TLI, bool HasBaseReg) {
771 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
772 int64_t VC = SC->getValue()->getSExtValue();
774 TargetLowering::AddrMode AM;
776 AM.HasBaseReg = HasBaseReg;
777 return TLI->isLegalAddressingMode(AM, UseTy);
779 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
780 return (VC > -(1 << 16) && VC < (1 << 16)-1);
784 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
785 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
787 TargetLowering::AddrMode AM;
789 AM.HasBaseReg = HasBaseReg;
790 return TLI->isLegalAddressingMode(AM, UseTy);
792 // Default: assume global addresses are not legal.
799 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
800 /// loop varying to the Imm operand.
801 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
802 Loop *L, ScalarEvolution *SE) {
803 if (Val->isLoopInvariant(L)) return; // Nothing to do.
805 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
806 std::vector<SCEVHandle> NewOps;
807 NewOps.reserve(SAE->getNumOperands());
809 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
810 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
811 // If this is a loop-variant expression, it must stay in the immediate
812 // field of the expression.
813 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
815 NewOps.push_back(SAE->getOperand(i));
819 Val = SE->getIntegerSCEV(0, Val->getType());
821 Val = SE->getAddExpr(NewOps);
822 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
823 // Try to pull immediates out of the start value of nested addrec's.
824 SCEVHandle Start = SARE->getStart();
825 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
827 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
829 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
831 // Otherwise, all of Val is variant, move the whole thing over.
832 Imm = SE->getAddExpr(Imm, Val);
833 Val = SE->getIntegerSCEV(0, Val->getType());
838 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
839 /// that can fit into the immediate field of instructions in the target.
840 /// Accumulate these immediate values into the Imm value.
841 static void MoveImmediateValues(const TargetLowering *TLI,
843 SCEVHandle &Val, SCEVHandle &Imm,
844 bool isAddress, Loop *L,
845 ScalarEvolution *SE) {
846 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
847 std::vector<SCEVHandle> NewOps;
848 NewOps.reserve(SAE->getNumOperands());
850 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
851 SCEVHandle NewOp = SAE->getOperand(i);
852 MoveImmediateValues(TLI, UseTy, NewOp, Imm, isAddress, L, SE);
854 if (!NewOp->isLoopInvariant(L)) {
855 // If this is a loop-variant expression, it must stay in the immediate
856 // field of the expression.
857 Imm = SE->getAddExpr(Imm, NewOp);
859 NewOps.push_back(NewOp);
864 Val = SE->getIntegerSCEV(0, Val->getType());
866 Val = SE->getAddExpr(NewOps);
868 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
869 // Try to pull immediates out of the start value of nested addrec's.
870 SCEVHandle Start = SARE->getStart();
871 MoveImmediateValues(TLI, UseTy, Start, Imm, isAddress, L, SE);
873 if (Start != SARE->getStart()) {
874 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
876 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
879 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
880 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
881 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
882 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
884 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
885 SCEVHandle NewOp = SME->getOperand(1);
886 MoveImmediateValues(TLI, UseTy, NewOp, SubImm, isAddress, L, SE);
888 // If we extracted something out of the subexpressions, see if we can
890 if (NewOp != SME->getOperand(1)) {
891 // Scale SubImm up by "8". If the result is a target constant, we are
893 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
894 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
895 // Accumulate the immediate.
896 Imm = SE->getAddExpr(Imm, SubImm);
898 // Update what is left of 'Val'.
899 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
906 // Loop-variant expressions must stay in the immediate field of the
908 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
909 !Val->isLoopInvariant(L)) {
910 Imm = SE->getAddExpr(Imm, Val);
911 Val = SE->getIntegerSCEV(0, Val->getType());
915 // Otherwise, no immediates to move.
918 static void MoveImmediateValues(const TargetLowering *TLI,
920 SCEVHandle &Val, SCEVHandle &Imm,
921 bool isAddress, Loop *L,
922 ScalarEvolution *SE) {
923 const Type *UseTy = getAccessType(User);
924 MoveImmediateValues(TLI, UseTy, Val, Imm, isAddress, L, SE);
927 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
928 /// added together. This is used to reassociate common addition subexprs
929 /// together for maximal sharing when rewriting bases.
930 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
932 ScalarEvolution *SE) {
933 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
934 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
935 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
936 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
937 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
938 if (SARE->getOperand(0) == Zero) {
939 SubExprs.push_back(Expr);
941 // Compute the addrec with zero as its base.
942 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
943 Ops[0] = Zero; // Start with zero base.
944 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
947 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
949 } else if (!Expr->isZero()) {
951 SubExprs.push_back(Expr);
955 // This is logically local to the following function, but C++ says we have
956 // to make it file scope.
957 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
959 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
960 /// the Uses, removing any common subexpressions, except that if all such
961 /// subexpressions can be folded into an addressing mode for all uses inside
962 /// the loop (this case is referred to as "free" in comments herein) we do
963 /// not remove anything. This looks for things like (a+b+c) and
964 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
965 /// is *removed* from the Bases and returned.
967 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
968 ScalarEvolution *SE, Loop *L,
969 const TargetLowering *TLI) {
970 unsigned NumUses = Uses.size();
972 // Only one use? This is a very common case, so we handle it specially and
974 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
975 SCEVHandle Result = Zero;
976 SCEVHandle FreeResult = Zero;
978 // If the use is inside the loop, use its base, regardless of what it is:
979 // it is clearly shared across all the IV's. If the use is outside the loop
980 // (which means after it) we don't want to factor anything *into* the loop,
981 // so just use 0 as the base.
982 if (L->contains(Uses[0].Inst->getParent()))
983 std::swap(Result, Uses[0].Base);
987 // To find common subexpressions, count how many of Uses use each expression.
988 // If any subexpressions are used Uses.size() times, they are common.
989 // Also track whether all uses of each expression can be moved into an
990 // an addressing mode "for free"; such expressions are left within the loop.
991 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
992 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
994 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
995 // order we see them.
996 std::vector<SCEVHandle> UniqueSubExprs;
998 std::vector<SCEVHandle> SubExprs;
999 unsigned NumUsesInsideLoop = 0;
1000 for (unsigned i = 0; i != NumUses; ++i) {
1001 // If the user is outside the loop, just ignore it for base computation.
1002 // Since the user is outside the loop, it must be *after* the loop (if it
1003 // were before, it could not be based on the loop IV). We don't want users
1004 // after the loop to affect base computation of values *inside* the loop,
1005 // because we can always add their offsets to the result IV after the loop
1006 // is done, ensuring we get good code inside the loop.
1007 if (!L->contains(Uses[i].Inst->getParent()))
1009 NumUsesInsideLoop++;
1011 // If the base is zero (which is common), return zero now, there are no
1012 // CSEs we can find.
1013 if (Uses[i].Base == Zero) return Zero;
1015 // If this use is as an address we may be able to put CSEs in the addressing
1016 // mode rather than hoisting them.
1017 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1018 // We may need the UseTy below, but only when isAddrUse, so compute it
1019 // only in that case.
1020 const Type *UseTy = 0;
1022 UseTy = getAccessType(Uses[i].Inst);
1024 // Split the expression into subexprs.
1025 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1026 // Add one to SubExpressionUseData.Count for each subexpr present, and
1027 // if the subexpr is not a valid immediate within an addressing mode use,
1028 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1029 // hoist these out of the loop (if they are common to all uses).
1030 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1031 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1032 UniqueSubExprs.push_back(SubExprs[j]);
1033 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1034 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1039 // Now that we know how many times each is used, build Result. Iterate over
1040 // UniqueSubexprs so that we have a stable ordering.
1041 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1042 std::map<SCEVHandle, SubExprUseData>::iterator I =
1043 SubExpressionUseData.find(UniqueSubExprs[i]);
1044 assert(I != SubExpressionUseData.end() && "Entry not found?");
1045 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1046 if (I->second.notAllUsesAreFree)
1047 Result = SE->getAddExpr(Result, I->first);
1049 FreeResult = SE->getAddExpr(FreeResult, I->first);
1051 // Remove non-cse's from SubExpressionUseData.
1052 SubExpressionUseData.erase(I);
1055 if (FreeResult != Zero) {
1056 // We have some subexpressions that can be subsumed into addressing
1057 // modes in every use inside the loop. However, it's possible that
1058 // there are so many of them that the combined FreeResult cannot
1059 // be subsumed, or that the target cannot handle both a FreeResult
1060 // and a Result in the same instruction (for example because it would
1061 // require too many registers). Check this.
1062 for (unsigned i=0; i<NumUses; ++i) {
1063 if (!L->contains(Uses[i].Inst->getParent()))
1065 // We know this is an addressing mode use; if there are any uses that
1066 // are not, FreeResult would be Zero.
1067 const Type *UseTy = getAccessType(Uses[i].Inst);
1068 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1069 // FIXME: could split up FreeResult into pieces here, some hoisted
1070 // and some not. There is no obvious advantage to this.
1071 Result = SE->getAddExpr(Result, FreeResult);
1078 // If we found no CSE's, return now.
1079 if (Result == Zero) return Result;
1081 // If we still have a FreeResult, remove its subexpressions from
1082 // SubExpressionUseData. This means they will remain in the use Bases.
1083 if (FreeResult != Zero) {
1084 SeparateSubExprs(SubExprs, FreeResult, SE);
1085 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1086 std::map<SCEVHandle, SubExprUseData>::iterator I =
1087 SubExpressionUseData.find(SubExprs[j]);
1088 SubExpressionUseData.erase(I);
1093 // Otherwise, remove all of the CSE's we found from each of the base values.
1094 for (unsigned i = 0; i != NumUses; ++i) {
1095 // Uses outside the loop don't necessarily include the common base, but
1096 // the final IV value coming into those uses does. Instead of trying to
1097 // remove the pieces of the common base, which might not be there,
1098 // subtract off the base to compensate for this.
1099 if (!L->contains(Uses[i].Inst->getParent())) {
1100 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1104 // Split the expression into subexprs.
1105 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1107 // Remove any common subexpressions.
1108 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1109 if (SubExpressionUseData.count(SubExprs[j])) {
1110 SubExprs.erase(SubExprs.begin()+j);
1114 // Finally, add the non-shared expressions together.
1115 if (SubExprs.empty())
1116 Uses[i].Base = Zero;
1118 Uses[i].Base = SE->getAddExpr(SubExprs);
1125 /// ValidStride - Check whether the given Scale is valid for all loads and
1126 /// stores in UsersToProcess.
1128 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1130 const std::vector<BasedUser>& UsersToProcess) {
1134 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1135 // If this is a load or other access, pass the type of the access in.
1136 const Type *AccessTy = Type::VoidTy;
1137 if (isAddressUse(UsersToProcess[i].Inst,
1138 UsersToProcess[i].OperandValToReplace))
1139 AccessTy = getAccessType(UsersToProcess[i].Inst);
1140 else if (isa<PHINode>(UsersToProcess[i].Inst))
1143 TargetLowering::AddrMode AM;
1144 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1145 AM.BaseOffs = SC->getValue()->getSExtValue();
1146 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1149 // If load[imm+r*scale] is illegal, bail out.
1150 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1156 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
1158 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1162 Ty1 = SE->getEffectiveSCEVType(Ty1);
1163 Ty2 = SE->getEffectiveSCEVType(Ty2);
1166 if (Ty1->canLosslesslyBitCastTo(Ty2))
1168 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1173 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1174 /// of a previous stride and it is a legal value for the target addressing
1175 /// mode scale component and optional base reg. This allows the users of
1176 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1177 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1179 /// If all uses are outside the loop, we don't require that all multiplies
1180 /// be folded into the addressing mode, nor even that the factor be constant;
1181 /// a multiply (executed once) outside the loop is better than another IV
1182 /// within. Well, usually.
1183 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1184 bool AllUsesAreAddresses,
1185 bool AllUsesAreOutsideLoop,
1186 const SCEVHandle &Stride,
1187 IVExpr &IV, const Type *Ty,
1188 const std::vector<BasedUser>& UsersToProcess) {
1189 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1190 int64_t SInt = SC->getValue()->getSExtValue();
1191 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1193 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1194 IVsByStride.find(StrideOrder[NewStride]);
1195 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1197 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1198 if (SI->first != Stride &&
1199 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1201 int64_t Scale = SInt / SSInt;
1202 // Check that this stride is valid for all the types used for loads and
1203 // stores; if it can be used for some and not others, we might as well use
1204 // the original stride everywhere, since we have to create the IV for it
1205 // anyway. If the scale is 1, then we don't need to worry about folding
1208 (AllUsesAreAddresses &&
1209 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1210 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1211 IE = SI->second.IVs.end(); II != IE; ++II)
1212 // FIXME: Only handle base == 0 for now.
1213 // Only reuse previous IV if it would not require a type conversion.
1214 if (II->Base->isZero() &&
1215 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1217 return SE->getIntegerSCEV(Scale, Stride->getType());
1220 } else if (AllUsesAreOutsideLoop) {
1221 // Accept nonconstant strides here; it is really really right to substitute
1222 // an existing IV if we can.
1223 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1225 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1226 IVsByStride.find(StrideOrder[NewStride]);
1227 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1229 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1230 if (SI->first != Stride && SSInt != 1)
1232 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1233 IE = SI->second.IVs.end(); II != IE; ++II)
1234 // Accept nonzero base here.
1235 // Only reuse previous IV if it would not require a type conversion.
1236 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1241 // Special case, old IV is -1*x and this one is x. Can treat this one as
1243 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1245 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1246 IVsByStride.find(StrideOrder[NewStride]);
1247 if (SI == IVsByStride.end())
1249 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1250 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1251 if (Stride == ME->getOperand(1) &&
1252 SC->getValue()->getSExtValue() == -1LL)
1253 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1254 IE = SI->second.IVs.end(); II != IE; ++II)
1255 // Accept nonzero base here.
1256 // Only reuse previous IV if it would not require type conversion.
1257 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1259 return SE->getIntegerSCEV(-1LL, Stride->getType());
1263 return SE->getIntegerSCEV(0, Stride->getType());
1266 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1267 /// returns true if Val's isUseOfPostIncrementedValue is true.
1268 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1269 return Val.isUseOfPostIncrementedValue;
1272 /// isNonConstantNegative - Return true if the specified scev is negated, but
1274 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1275 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1276 if (!Mul) return false;
1278 // If there is a constant factor, it will be first.
1279 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1280 if (!SC) return false;
1282 // Return true if the value is negative, this matches things like (-42 * V).
1283 return SC->getValue()->getValue().isNegative();
1286 // CollectIVUsers - Transform our list of users and offsets to a bit more
1287 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1288 // of the strided accesses, as well as the old information from Uses. We
1289 // progressively move information from the Base field to the Imm field, until
1290 // we eventually have the full access expression to rewrite the use.
1291 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1292 IVUsersOfOneStride &Uses,
1294 bool &AllUsesAreAddresses,
1295 bool &AllUsesAreOutsideLoop,
1296 std::vector<BasedUser> &UsersToProcess) {
1297 UsersToProcess.reserve(Uses.Users.size());
1298 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1299 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1301 // Move any loop variant operands from the offset field to the immediate
1302 // field of the use, so that we don't try to use something before it is
1304 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1305 UsersToProcess.back().Imm, L, SE);
1306 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1307 "Base value is not loop invariant!");
1310 // We now have a whole bunch of uses of like-strided induction variables, but
1311 // they might all have different bases. We want to emit one PHI node for this
1312 // stride which we fold as many common expressions (between the IVs) into as
1313 // possible. Start by identifying the common expressions in the base values
1314 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1315 // "A+B"), emit it to the preheader, then remove the expression from the
1316 // UsersToProcess base values.
1317 SCEVHandle CommonExprs =
1318 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1320 // Next, figure out what we can represent in the immediate fields of
1321 // instructions. If we can represent anything there, move it to the imm
1322 // fields of the BasedUsers. We do this so that it increases the commonality
1323 // of the remaining uses.
1324 unsigned NumPHI = 0;
1325 bool HasAddress = false;
1326 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1327 // If the user is not in the current loop, this means it is using the exit
1328 // value of the IV. Do not put anything in the base, make sure it's all in
1329 // the immediate field to allow as much factoring as possible.
1330 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1331 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1332 UsersToProcess[i].Base);
1333 UsersToProcess[i].Base =
1334 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1336 // Not all uses are outside the loop.
1337 AllUsesAreOutsideLoop = false;
1339 // Addressing modes can be folded into loads and stores. Be careful that
1340 // the store is through the expression, not of the expression though.
1342 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1343 UsersToProcess[i].OperandValToReplace);
1344 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1352 // If this use isn't an address, then not all uses are addresses.
1353 if (!isAddress && !isPHI)
1354 AllUsesAreAddresses = false;
1356 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1357 UsersToProcess[i].Imm, isAddress, L, SE);
1361 // If one of the use is a PHI node and all other uses are addresses, still
1362 // allow iv reuse. Essentially we are trading one constant multiplication
1363 // for one fewer iv.
1365 AllUsesAreAddresses = false;
1367 // There are no in-loop address uses.
1368 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1369 AllUsesAreAddresses = false;
1374 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1375 /// is valid and profitable for the given set of users of a stride. In
1376 /// full strength-reduction mode, all addresses at the current stride are
1377 /// strength-reduced all the way down to pointer arithmetic.
1379 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1380 const std::vector<BasedUser> &UsersToProcess,
1382 bool AllUsesAreAddresses,
1383 SCEVHandle Stride) {
1384 if (!EnableFullLSRMode)
1387 // The heuristics below aim to avoid increasing register pressure, but
1388 // fully strength-reducing all the addresses increases the number of
1389 // add instructions, so don't do this when optimizing for size.
1390 // TODO: If the loop is large, the savings due to simpler addresses
1391 // may oughtweight the costs of the extra increment instructions.
1392 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1395 // TODO: For now, don't do full strength reduction if there could
1396 // potentially be greater-stride multiples of the current stride
1397 // which could reuse the current stride IV.
1398 if (StrideOrder.back() != Stride)
1401 // Iterate through the uses to find conditions that automatically rule out
1403 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1404 const SCEV *Base = UsersToProcess[i].Base;
1405 const SCEV *Imm = UsersToProcess[i].Imm;
1406 // If any users have a loop-variant component, they can't be fully
1407 // strength-reduced.
1408 if (Imm && !Imm->isLoopInvariant(L))
1410 // If there are to users with the same base and the difference between
1411 // the two Imm values can't be folded into the address, full
1412 // strength reduction would increase register pressure.
1414 const SCEV *CurImm = UsersToProcess[i].Imm;
1415 if ((CurImm || Imm) && CurImm != Imm) {
1416 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1417 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1418 const Instruction *Inst = UsersToProcess[i].Inst;
1419 const Type *UseTy = getAccessType(Inst);
1420 SCEVHandle Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1421 if (!Diff->isZero() &&
1422 (!AllUsesAreAddresses ||
1423 !fitsInAddressMode(Diff, UseTy, TLI, /*HasBaseReg=*/true)))
1426 } while (++i != e && Base == UsersToProcess[i].Base);
1429 // If there's exactly one user in this stride, fully strength-reducing it
1430 // won't increase register pressure. If it's starting from a non-zero base,
1431 // it'll be simpler this way.
1432 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1435 // Otherwise, if there are any users in this stride that don't require
1436 // a register for their base, full strength-reduction will increase
1437 // register pressure.
1438 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1439 if (UsersToProcess[i].Base->isZero())
1442 // Otherwise, go for it.
1446 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1447 /// with the specified start and step values in the specified loop.
1449 /// If NegateStride is true, the stride should be negated by using a
1450 /// subtract instead of an add.
1452 /// Return the created phi node.
1454 static PHINode *InsertAffinePhi(SCEVHandle Start, SCEVHandle Step,
1456 SCEVExpander &Rewriter) {
1457 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1458 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1460 BasicBlock *Header = L->getHeader();
1461 BasicBlock *Preheader = L->getLoopPreheader();
1462 BasicBlock *LatchBlock = L->getLoopLatch();
1463 const Type *Ty = Start->getType();
1464 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1466 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1467 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1470 // If the stride is negative, insert a sub instead of an add for the
1472 bool isNegative = isNonConstantNegative(Step);
1473 SCEVHandle IncAmount = Step;
1475 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1477 // Insert an add instruction right before the terminator corresponding
1478 // to the back-edge.
1479 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1480 Preheader->getTerminator());
1483 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1484 LatchBlock->getTerminator());
1486 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1487 LatchBlock->getTerminator());
1489 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1491 PN->addIncoming(IncV, LatchBlock);
1497 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1498 // We want to emit code for users inside the loop first. To do this, we
1499 // rearrange BasedUser so that the entries at the end have
1500 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1501 // vector (so we handle them first).
1502 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1503 PartitionByIsUseOfPostIncrementedValue);
1505 // Sort this by base, so that things with the same base are handled
1506 // together. By partitioning first and stable-sorting later, we are
1507 // guaranteed that within each base we will pop off users from within the
1508 // loop before users outside of the loop with a particular base.
1510 // We would like to use stable_sort here, but we can't. The problem is that
1511 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1512 // we don't have anything to do a '<' comparison on. Because we think the
1513 // number of uses is small, do a horrible bubble sort which just relies on
1515 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1516 // Get a base value.
1517 SCEVHandle Base = UsersToProcess[i].Base;
1519 // Compact everything with this base to be consecutive with this one.
1520 for (unsigned j = i+1; j != e; ++j) {
1521 if (UsersToProcess[j].Base == Base) {
1522 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1529 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1530 /// UsersToProcess, meaning lowering addresses all the way down to direct
1531 /// pointer arithmetic.
1534 LoopStrengthReduce::PrepareToStrengthReduceFully(
1535 std::vector<BasedUser> &UsersToProcess,
1537 SCEVHandle CommonExprs,
1539 SCEVExpander &PreheaderRewriter) {
1540 DOUT << " Fully reducing all users\n";
1542 // Rewrite the UsersToProcess records, creating a separate PHI for each
1543 // unique Base value.
1544 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1545 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1546 // pick the first Imm value here to start with, and adjust it for the
1548 SCEVHandle Imm = UsersToProcess[i].Imm;
1549 SCEVHandle Base = UsersToProcess[i].Base;
1550 SCEVHandle Start = SE->getAddExpr(CommonExprs, Base, Imm);
1551 PHINode *Phi = InsertAffinePhi(Start, Stride, L,
1553 // Loop over all the users with the same base.
1555 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1556 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1557 UsersToProcess[i].Phi = Phi;
1558 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1559 "ShouldUseFullStrengthReductionMode should reject this!");
1560 } while (++i != e && Base == UsersToProcess[i].Base);
1564 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1565 /// given users to share.
1568 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1569 std::vector<BasedUser> &UsersToProcess,
1571 SCEVHandle CommonExprs,
1574 SCEVExpander &PreheaderRewriter) {
1575 DOUT << " Inserting new PHI:\n";
1577 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1581 // Remember this in case a later stride is multiple of this.
1582 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1584 // All the users will share this new IV.
1585 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1586 UsersToProcess[i].Phi = Phi;
1589 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1593 /// PrepareToStrengthReduceWithNewPhi - Prepare for the given users to reuse
1594 /// an induction variable with a stride that is a factor of the current
1595 /// induction variable.
1598 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1599 std::vector<BasedUser> &UsersToProcess,
1601 const IVExpr &ReuseIV,
1602 Instruction *PreInsertPt) {
1603 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1604 << " and BASE " << *ReuseIV.Base << "\n";
1606 // All the users will share the reused IV.
1607 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1608 UsersToProcess[i].Phi = ReuseIV.PHI;
1610 Constant *C = dyn_cast<Constant>(CommonBaseV);
1612 (!C->isNullValue() &&
1613 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1615 // We want the common base emitted into the preheader! This is just
1616 // using cast as a copy so BitCast (no-op cast) is appropriate
1617 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1618 "commonbase", PreInsertPt);
1621 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1622 const Type *AccessTy,
1623 std::vector<BasedUser> &UsersToProcess,
1624 const TargetLowering *TLI) {
1625 SmallVector<Instruction*, 16> AddrModeInsts;
1626 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1627 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1629 ExtAddrMode AddrMode =
1630 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1631 AccessTy, UsersToProcess[i].Inst,
1632 AddrModeInsts, *TLI);
1633 if (GV && GV != AddrMode.BaseGV)
1635 if (Offset && !AddrMode.BaseOffs)
1636 // FIXME: How to accurate check it's immediate offset is folded.
1638 AddrModeInsts.clear();
1643 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1644 /// stride of IV. All of the users may have different starting values, and this
1645 /// may not be the only stride.
1646 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1647 IVUsersOfOneStride &Uses,
1649 // If all the users are moved to another stride, then there is nothing to do.
1650 if (Uses.Users.empty())
1653 // Keep track if every use in UsersToProcess is an address. If they all are,
1654 // we may be able to rewrite the entire collection of them in terms of a
1655 // smaller-stride IV.
1656 bool AllUsesAreAddresses = true;
1658 // Keep track if every use of a single stride is outside the loop. If so,
1659 // we want to be more aggressive about reusing a smaller-stride IV; a
1660 // multiply outside the loop is better than another IV inside. Well, usually.
1661 bool AllUsesAreOutsideLoop = true;
1663 // Transform our list of users and offsets to a bit more complex table. In
1664 // this new vector, each 'BasedUser' contains 'Base' the base of the
1665 // strided accessas well as the old information from Uses. We progressively
1666 // move information from the Base field to the Imm field, until we eventually
1667 // have the full access expression to rewrite the use.
1668 std::vector<BasedUser> UsersToProcess;
1669 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1670 AllUsesAreOutsideLoop,
1673 // Sort the UsersToProcess array so that users with common bases are
1674 // next to each other.
1675 SortUsersToProcess(UsersToProcess);
1677 // If we managed to find some expressions in common, we'll need to carry
1678 // their value in a register and add it in for each use. This will take up
1679 // a register operand, which potentially restricts what stride values are
1681 bool HaveCommonExprs = !CommonExprs->isZero();
1682 const Type *ReplacedTy = CommonExprs->getType();
1684 // If all uses are addresses, consider sinking the immediate part of the
1685 // common expression back into uses if they can fit in the immediate fields.
1686 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1687 SCEVHandle NewCommon = CommonExprs;
1688 SCEVHandle Imm = SE->getIntegerSCEV(0, ReplacedTy);
1689 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE);
1690 if (!Imm->isZero()) {
1693 // If the immediate part of the common expression is a GV, check if it's
1694 // possible to fold it into the target addressing mode.
1695 GlobalValue *GV = 0;
1696 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1697 GV = dyn_cast<GlobalValue>(SU->getValue());
1699 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1700 Offset = SC->getValue()->getSExtValue();
1702 // Pass VoidTy as the AccessTy to be conservative, because
1703 // there could be multiple access types among all the uses.
1704 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy,
1705 UsersToProcess, TLI);
1708 DOUT << " Sinking " << *Imm << " back down into uses\n";
1709 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1710 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1711 CommonExprs = NewCommon;
1712 HaveCommonExprs = !CommonExprs->isZero();
1718 // Now that we know what we need to do, insert the PHI node itself.
1720 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1722 << " Common base: " << *CommonExprs << "\n";
1724 SCEVExpander Rewriter(*SE, *LI);
1725 SCEVExpander PreheaderRewriter(*SE, *LI);
1727 BasicBlock *Preheader = L->getLoopPreheader();
1728 Instruction *PreInsertPt = Preheader->getTerminator();
1729 BasicBlock *LatchBlock = L->getLoopLatch();
1731 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1733 SCEVHandle RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1734 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1735 SE->getIntegerSCEV(0, Type::Int32Ty),
1738 /// Choose a strength-reduction strategy and prepare for it by creating
1739 /// the necessary PHIs and adjusting the bookkeeping.
1740 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1741 AllUsesAreAddresses, Stride)) {
1742 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1745 // Emit the initial base value into the loop preheader.
1746 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1749 // If all uses are addresses, check if it is possible to reuse an IV. The
1750 // new IV must have a stride that is a multiple of the old stride; the
1751 // multiple must be a number that can be encoded in the scale field of the
1752 // target addressing mode; and we must have a valid instruction after this
1753 // substitution, including the immediate field, if any.
1754 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1755 AllUsesAreOutsideLoop,
1756 Stride, ReuseIV, ReplacedTy,
1758 if (isa<SCEVConstant>(RewriteFactor) &&
1759 cast<SCEVConstant>(RewriteFactor)->isZero())
1760 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1761 CommonBaseV, L, PreheaderRewriter);
1763 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1764 ReuseIV, PreInsertPt);
1767 // Process all the users now, replacing their strided uses with
1768 // strength-reduced forms. This outer loop handles all bases, the inner
1769 // loop handles all users of a particular base.
1770 while (!UsersToProcess.empty()) {
1771 SCEVHandle Base = UsersToProcess.back().Base;
1772 Instruction *Inst = UsersToProcess.back().Inst;
1774 // Emit the code for Base into the preheader.
1776 if (!Base->isZero()) {
1777 BaseV = PreheaderRewriter.expandCodeFor(Base, Base->getType(),
1780 DOUT << " INSERTING code for BASE = " << *Base << ":";
1781 if (BaseV->hasName())
1782 DOUT << " Result value name = %" << BaseV->getNameStr();
1785 // If BaseV is a non-zero constant, make sure that it gets inserted into
1786 // the preheader, instead of being forward substituted into the uses. We
1787 // do this by forcing a BitCast (noop cast) to be inserted into the
1788 // preheader in this case.
1789 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false)) {
1790 // We want this constant emitted into the preheader! This is just
1791 // using cast as a copy so BitCast (no-op cast) is appropriate
1792 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1797 // Emit the code to add the immediate offset to the Phi value, just before
1798 // the instructions that we identified as using this stride and base.
1800 // FIXME: Use emitted users to emit other users.
1801 BasedUser &User = UsersToProcess.back();
1803 DOUT << " Examining use ";
1804 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1805 /*PrintType=*/false));
1806 DOUT << " in Inst: " << *(User.Inst);
1808 // If this instruction wants to use the post-incremented value, move it
1809 // after the post-inc and use its value instead of the PHI.
1810 Value *RewriteOp = User.Phi;
1811 if (User.isUseOfPostIncrementedValue) {
1812 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1814 // If this user is in the loop, make sure it is the last thing in the
1815 // loop to ensure it is dominated by the increment.
1816 if (L->contains(User.Inst->getParent()))
1817 User.Inst->moveBefore(LatchBlock->getTerminator());
1820 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1822 if (SE->getTypeSizeInBits(RewriteOp->getType()) !=
1823 SE->getTypeSizeInBits(ReplacedTy)) {
1824 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1825 SE->getTypeSizeInBits(ReplacedTy) &&
1826 "Unexpected widening cast!");
1827 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1830 // If we had to insert new instructions for RewriteOp, we have to
1831 // consider that they may not have been able to end up immediately
1832 // next to RewriteOp, because non-PHI instructions may never precede
1833 // PHI instructions in a block. In this case, remember where the last
1834 // instruction was inserted so that if we're replacing a different
1835 // PHI node, we can use the later point to expand the final
1837 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1838 if (RewriteOp == User.Phi) NewBasePt = 0;
1840 // Clear the SCEVExpander's expression map so that we are guaranteed
1841 // to have the code emitted where we expect it.
1844 // If we are reusing the iv, then it must be multiplied by a constant
1845 // factor to take advantage of the addressing mode scale component.
1846 if (!RewriteFactor->isZero()) {
1847 // If we're reusing an IV with a nonzero base (currently this happens
1848 // only when all reuses are outside the loop) subtract that base here.
1849 // The base has been used to initialize the PHI node but we don't want
1851 if (!ReuseIV.Base->isZero()) {
1852 SCEVHandle typedBase = ReuseIV.Base;
1853 if (SE->getTypeSizeInBits(RewriteExpr->getType()) !=
1854 SE->getTypeSizeInBits(ReuseIV.Base->getType())) {
1855 // It's possible the original IV is a larger type than the new IV,
1856 // in which case we have to truncate the Base. We checked in
1857 // RequiresTypeConversion that this is valid.
1858 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1859 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1860 "Unexpected lengthening conversion!");
1861 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1862 RewriteExpr->getType());
1864 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1867 // Multiply old variable, with base removed, by new scale factor.
1868 RewriteExpr = SE->getMulExpr(RewriteFactor,
1871 // The common base is emitted in the loop preheader. But since we
1872 // are reusing an IV, it has not been used to initialize the PHI node.
1873 // Add it to the expression used to rewrite the uses.
1874 // When this use is outside the loop, we earlier subtracted the
1875 // common base, and are adding it back here. Use the same expression
1876 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1877 if (!CommonExprs->isZero()) {
1878 if (L->contains(User.Inst->getParent()))
1879 RewriteExpr = SE->getAddExpr(RewriteExpr,
1880 SE->getUnknown(CommonBaseV));
1882 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1886 // Now that we know what we need to do, insert code before User for the
1887 // immediate and any loop-variant expressions.
1889 // Add BaseV to the PHI value if needed.
1890 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1892 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1896 // Mark old value we replaced as possibly dead, so that it is eliminated
1897 // if we just replaced the last use of that value.
1898 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1900 UsersToProcess.pop_back();
1903 // If there are any more users to process with the same base, process them
1904 // now. We sorted by base above, so we just have to check the last elt.
1905 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1906 // TODO: Next, find out which base index is the most common, pull it out.
1909 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1910 // different starting values, into different PHIs.
1913 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1914 /// set the IV user and stride information and return true, otherwise return
1916 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1917 const SCEVHandle *&CondStride) {
1918 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1920 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1921 IVUsesByStride.find(StrideOrder[Stride]);
1922 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1924 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1925 E = SI->second.Users.end(); UI != E; ++UI)
1926 if (UI->User == Cond) {
1927 // NOTE: we could handle setcc instructions with multiple uses here, but
1928 // InstCombine does it as well for simple uses, it's not clear that it
1929 // occurs enough in real life to handle.
1931 CondStride = &SI->first;
1939 // Constant strides come first which in turns are sorted by their absolute
1940 // values. If absolute values are the same, then positive strides comes first.
1942 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1943 struct StrideCompare {
1944 const ScalarEvolution *SE;
1945 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1947 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1948 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1949 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1951 int64_t LV = LHSC->getValue()->getSExtValue();
1952 int64_t RV = RHSC->getValue()->getSExtValue();
1953 uint64_t ALV = (LV < 0) ? -LV : LV;
1954 uint64_t ARV = (RV < 0) ? -RV : RV;
1962 // If it's the same value but different type, sort by bit width so
1963 // that we emit larger induction variables before smaller
1964 // ones, letting the smaller be re-written in terms of larger ones.
1965 return SE->getTypeSizeInBits(RHS->getType()) <
1966 SE->getTypeSizeInBits(LHS->getType());
1968 return LHSC && !RHSC;
1973 /// ChangeCompareStride - If a loop termination compare instruction is the
1974 /// only use of its stride, and the compaison is against a constant value,
1975 /// try eliminate the stride by moving the compare instruction to another
1976 /// stride and change its constant operand accordingly. e.g.
1982 /// if (v2 < 10) goto loop
1987 /// if (v1 < 30) goto loop
1988 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1989 IVStrideUse* &CondUse,
1990 const SCEVHandle* &CondStride) {
1991 if (StrideOrder.size() < 2 ||
1992 IVUsesByStride[*CondStride].Users.size() != 1)
1994 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1995 if (!SC) return Cond;
1997 ICmpInst::Predicate Predicate = Cond->getPredicate();
1998 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1999 unsigned BitWidth = SE->getTypeSizeInBits((*CondStride)->getType());
2000 uint64_t SignBit = 1ULL << (BitWidth-1);
2001 const Type *CmpTy = Cond->getOperand(0)->getType();
2002 const Type *NewCmpTy = NULL;
2003 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
2004 unsigned NewTyBits = 0;
2005 SCEVHandle *NewStride = NULL;
2006 Value *NewCmpLHS = NULL;
2007 Value *NewCmpRHS = NULL;
2009 SCEVHandle NewOffset = SE->getIntegerSCEV(0, CmpTy);
2011 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
2012 int64_t CmpVal = C->getValue().getSExtValue();
2014 // Check stride constant and the comparision constant signs to detect
2016 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
2019 // Look for a suitable stride / iv as replacement.
2020 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
2021 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2022 IVUsesByStride.find(StrideOrder[i]);
2023 if (!isa<SCEVConstant>(SI->first))
2025 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2026 if (SSInt == CmpSSInt ||
2027 abs(SSInt) < abs(CmpSSInt) ||
2028 (SSInt % CmpSSInt) != 0)
2031 Scale = SSInt / CmpSSInt;
2032 int64_t NewCmpVal = CmpVal * Scale;
2033 APInt Mul = APInt(BitWidth*2, CmpVal, true);
2034 Mul = Mul * APInt(BitWidth*2, Scale, true);
2035 // Check for overflow.
2036 if (!Mul.isSignedIntN(BitWidth))
2039 // Watch out for overflow.
2040 if (ICmpInst::isSignedPredicate(Predicate) &&
2041 (CmpVal & SignBit) != (NewCmpVal & SignBit))
2044 if (NewCmpVal == CmpVal)
2046 // Pick the best iv to use trying to avoid a cast.
2048 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2049 E = SI->second.Users.end(); UI != E; ++UI) {
2050 NewCmpLHS = UI->OperandValToReplace;
2051 if (NewCmpLHS->getType() == CmpTy)
2057 NewCmpTy = NewCmpLHS->getType();
2058 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
2059 const Type *NewCmpIntTy = IntegerType::get(NewTyBits);
2060 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2061 // Check if it is possible to rewrite it using
2062 // an iv / stride of a smaller integer type.
2063 unsigned Bits = NewTyBits;
2064 if (ICmpInst::isSignedPredicate(Predicate))
2066 uint64_t Mask = (1ULL << Bits) - 1;
2067 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2071 // Don't rewrite if use offset is non-constant and the new type is
2072 // of a different type.
2073 // FIXME: too conservative?
2074 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset))
2077 bool AllUsesAreAddresses = true;
2078 bool AllUsesAreOutsideLoop = true;
2079 std::vector<BasedUser> UsersToProcess;
2080 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
2081 AllUsesAreAddresses,
2082 AllUsesAreOutsideLoop,
2084 // Avoid rewriting the compare instruction with an iv of new stride
2085 // if it's likely the new stride uses will be rewritten using the
2086 // stride of the compare instruction.
2087 if (AllUsesAreAddresses &&
2088 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess))
2091 // If scale is negative, use swapped predicate unless it's testing
2093 if (Scale < 0 && !Cond->isEquality())
2094 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2096 NewStride = &StrideOrder[i];
2097 if (!isa<PointerType>(NewCmpTy))
2098 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2100 ConstantInt *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2101 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2103 NewOffset = TyBits == NewTyBits
2104 ? SE->getMulExpr(CondUse->Offset,
2105 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
2106 : SE->getConstant(ConstantInt::get(NewCmpIntTy,
2107 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
2112 // Forgo this transformation if it the increment happens to be
2113 // unfortunately positioned after the condition, and the condition
2114 // has multiple uses which prevent it from being moved immediately
2115 // before the branch. See
2116 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2117 // for an example of this situation.
2118 if (!Cond->hasOneUse()) {
2119 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2126 // Create a new compare instruction using new stride / iv.
2127 ICmpInst *OldCond = Cond;
2128 // Insert new compare instruction.
2129 Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
2130 L->getHeader()->getName() + ".termcond",
2133 // Remove the old compare instruction. The old indvar is probably dead too.
2134 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
2135 OldCond->replaceAllUsesWith(Cond);
2136 OldCond->eraseFromParent();
2138 IVUsesByStride[*CondStride].Users.pop_back();
2139 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewCmpLHS);
2140 CondUse = &IVUsesByStride[*NewStride].Users.back();
2141 CondStride = NewStride;
2149 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
2150 /// an smax computation.
2152 /// This is a narrow solution to a specific, but acute, problem. For loops
2158 /// } while (++i < n);
2160 /// where the comparison is signed, the trip count isn't just 'n', because
2161 /// 'n' could be negative. And unfortunately this can come up even for loops
2162 /// where the user didn't use a C do-while loop. For example, seemingly
2163 /// well-behaved top-test loops will commonly be lowered like this:
2169 /// } while (++i < n);
2172 /// and then it's possible for subsequent optimization to obscure the if
2173 /// test in such a way that indvars can't find it.
2175 /// When indvars can't find the if test in loops like this, it creates a
2176 /// signed-max expression, which allows it to give the loop a canonical
2177 /// induction variable:
2180 /// smax = n < 1 ? 1 : n;
2183 /// } while (++i != smax);
2185 /// Canonical induction variables are necessary because the loop passes
2186 /// are designed around them. The most obvious example of this is the
2187 /// LoopInfo analysis, which doesn't remember trip count values. It
2188 /// expects to be able to rediscover the trip count each time it is
2189 /// needed, and it does this using a simple analyis that only succeeds if
2190 /// the loop has a canonical induction variable.
2192 /// However, when it comes time to generate code, the maximum operation
2193 /// can be quite costly, especially if it's inside of an outer loop.
2195 /// This function solves this problem by detecting this type of loop and
2196 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2197 /// the instructions for the maximum computation.
2199 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
2200 IVStrideUse* &CondUse) {
2201 // Check that the loop matches the pattern we're looking for.
2202 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2203 Cond->getPredicate() != CmpInst::ICMP_NE)
2206 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2207 if (!Sel || !Sel->hasOneUse()) return Cond;
2209 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2210 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2212 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2214 // Add one to the backedge-taken count to get the trip count.
2215 SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2217 // Check for a max calculation that matches the pattern.
2218 const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
2219 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
2221 SCEVHandle SMaxLHS = SMax->getOperand(0);
2222 SCEVHandle SMaxRHS = SMax->getOperand(1);
2223 if (!SMaxLHS || SMaxLHS != One) return Cond;
2225 // Check the relevant induction variable for conformance to
2227 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2228 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2229 if (!AR || !AR->isAffine() ||
2230 AR->getStart() != One ||
2231 AR->getStepRecurrence(*SE) != One)
2234 assert(AR->getLoop() == L &&
2235 "Loop condition operand is an addrec in a different loop!");
2237 // Check the right operand of the select, and remember it, as it will
2238 // be used in the new comparison instruction.
2240 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
2241 NewRHS = Sel->getOperand(1);
2242 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
2243 NewRHS = Sel->getOperand(2);
2244 if (!NewRHS) return Cond;
2246 // Ok, everything looks ok to change the condition into an SLT or SGE and
2247 // delete the max calculation.
2249 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2252 Cond->getOperand(0), NewRHS, "scmp", Cond);
2254 // Delete the max calculation instructions.
2255 Cond->replaceAllUsesWith(NewCond);
2256 Cond->eraseFromParent();
2257 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2258 Sel->eraseFromParent();
2259 if (Cmp->use_empty())
2260 Cmp->eraseFromParent();
2261 CondUse->User = NewCond;
2265 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2266 /// inside the loop then try to eliminate the cast opeation.
2267 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2269 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2270 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2273 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2275 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2276 IVUsesByStride.find(StrideOrder[Stride]);
2277 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2278 if (!isa<SCEVConstant>(SI->first))
2281 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2282 E = SI->second.Users.end(); UI != E; /* empty */) {
2283 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2285 Instruction *ShadowUse = CandidateUI->User;
2286 const Type *DestTy = NULL;
2288 /* If shadow use is a int->float cast then insert a second IV
2289 to eliminate this cast.
2291 for (unsigned i = 0; i < n; ++i)
2297 for (unsigned i = 0; i < n; ++i, ++d)
2300 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2301 DestTy = UCast->getDestTy();
2302 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2303 DestTy = SCast->getDestTy();
2304 if (!DestTy) continue;
2307 /* If target does not support DestTy natively then do not apply
2308 this transformation. */
2309 MVT DVT = TLI->getValueType(DestTy);
2310 if (!TLI->isTypeLegal(DVT)) continue;
2313 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2315 if (PH->getNumIncomingValues() != 2) continue;
2317 const Type *SrcTy = PH->getType();
2318 int Mantissa = DestTy->getFPMantissaWidth();
2319 if (Mantissa == -1) continue;
2320 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2323 unsigned Entry, Latch;
2324 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2332 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2333 if (!Init) continue;
2334 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2336 BinaryOperator *Incr =
2337 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2338 if (!Incr) continue;
2339 if (Incr->getOpcode() != Instruction::Add
2340 && Incr->getOpcode() != Instruction::Sub)
2343 /* Initialize new IV, double d = 0.0 in above example. */
2344 ConstantInt *C = NULL;
2345 if (Incr->getOperand(0) == PH)
2346 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2347 else if (Incr->getOperand(1) == PH)
2348 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2354 /* Add new PHINode. */
2355 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2357 /* create new increment. '++d' in above example. */
2358 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2359 BinaryOperator *NewIncr =
2360 BinaryOperator::Create(Incr->getOpcode(),
2361 NewPH, CFP, "IV.S.next.", Incr);
2363 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2364 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2366 /* Remove cast operation */
2367 ShadowUse->replaceAllUsesWith(NewPH);
2368 ShadowUse->eraseFromParent();
2369 SI->second.Users.erase(CandidateUI);
2376 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2377 // uses in the loop, look to see if we can eliminate some, in favor of using
2378 // common indvars for the different uses.
2379 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2380 // TODO: implement optzns here.
2382 OptimizeShadowIV(L);
2384 OptimizeLoopTermCond(L);
2387 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2388 /// postinc iv when possible.
2389 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2390 // Finally, get the terminating condition for the loop if possible. If we
2391 // can, we want to change it to use a post-incremented version of its
2392 // induction variable, to allow coalescing the live ranges for the IV into
2393 // one register value.
2394 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2395 BasicBlock *Preheader = L->getLoopPreheader();
2396 BasicBlock *LatchBlock =
2397 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2398 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2399 if (!TermBr || TermBr->isUnconditional() ||
2400 !isa<ICmpInst>(TermBr->getCondition()))
2402 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2404 // Search IVUsesByStride to find Cond's IVUse if there is one.
2405 IVStrideUse *CondUse = 0;
2406 const SCEVHandle *CondStride = 0;
2408 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2409 return; // setcc doesn't use the IV.
2411 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2412 // being unable to find a sufficient guard, for example), change the loop
2413 // comparison to use SLT instead of NE.
2414 Cond = OptimizeSMax(L, Cond, CondUse);
2416 // If possible, change stride and operands of the compare instruction to
2417 // eliminate one stride.
2418 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2420 // It's possible for the setcc instruction to be anywhere in the loop, and
2421 // possible for it to have multiple users. If it is not immediately before
2422 // the latch block branch, move it.
2423 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2424 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2425 Cond->moveBefore(TermBr);
2427 // Otherwise, clone the terminating condition and insert into the loopend.
2428 Cond = cast<ICmpInst>(Cond->clone());
2429 Cond->setName(L->getHeader()->getName() + ".termcond");
2430 LatchBlock->getInstList().insert(TermBr, Cond);
2432 // Clone the IVUse, as the old use still exists!
2433 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2434 CondUse->OperandValToReplace);
2435 CondUse = &IVUsesByStride[*CondStride].Users.back();
2439 // If we get to here, we know that we can transform the setcc instruction to
2440 // use the post-incremented version of the IV, allowing us to coalesce the
2441 // live ranges for the IV correctly.
2442 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2443 CondUse->isUseOfPostIncrementedValue = true;
2447 // OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2448 // when to exit the loop is used only for that purpose, try to rearrange things
2449 // so it counts down to a test against zero.
2450 void LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2452 // If the number of times the loop is executed isn't computable, give up.
2453 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2454 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2457 // Get the terminating condition for the loop if possible (this isn't
2458 // necessarily in the latch, or a block that's a predecessor of the header).
2459 SmallVector<BasicBlock*, 8> ExitBlocks;
2460 L->getExitBlocks(ExitBlocks);
2461 if (ExitBlocks.size() != 1) return;
2463 // Okay, there is one exit block. Try to find the condition that causes the
2464 // loop to be exited.
2465 BasicBlock *ExitBlock = ExitBlocks[0];
2467 BasicBlock *ExitingBlock = 0;
2468 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
2470 if (L->contains(*PI)) {
2471 if (ExitingBlock == 0)
2474 return; // More than one block exiting!
2476 assert(ExitingBlock && "No exits from loop, something is broken!");
2478 // Okay, we've computed the exiting block. See what condition causes us to
2481 // FIXME: we should be able to handle switch instructions (with a single exit)
2482 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2483 if (TermBr == 0) return;
2484 assert(TermBr->isConditional() && "If unconditional, it can't be in loop!");
2485 if (!isa<ICmpInst>(TermBr->getCondition()))
2487 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2489 // Handle only tests for equality for the moment, and only stride 1.
2490 if (Cond->getPredicate() != CmpInst::ICMP_EQ)
2492 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2493 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2494 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2495 if (!AR || !AR->isAffine() || AR->getStepRecurrence(*SE) != One)
2498 // Make sure the IV is only used for counting. Value may be preinc or
2499 // postinc; 2 uses in either case.
2500 if (!Cond->getOperand(0)->hasNUses(2))
2502 PHINode *phi = dyn_cast<PHINode>(Cond->getOperand(0));
2504 if (phi && phi->getParent()==L->getHeader()) {
2505 // value tested is preinc. Find the increment.
2506 // A CmpInst is not a BinaryOperator; we depend on this.
2507 Instruction::use_iterator UI = phi->use_begin();
2508 incr = dyn_cast<BinaryOperator>(UI);
2510 incr = dyn_cast<BinaryOperator>(++UI);
2511 // 1 use for postinc value, the phi. Unnecessarily conservative?
2512 if (!incr || !incr->hasOneUse() || incr->getOpcode()!=Instruction::Add)
2515 // Value tested is postinc. Find the phi node.
2516 incr = dyn_cast<BinaryOperator>(Cond->getOperand(0));
2517 if (!incr || incr->getOpcode()!=Instruction::Add)
2520 Instruction::use_iterator UI = Cond->getOperand(0)->use_begin();
2521 phi = dyn_cast<PHINode>(UI);
2523 phi = dyn_cast<PHINode>(++UI);
2524 // 1 use for preinc value, the increment.
2525 if (!phi || phi->getParent()!=L->getHeader() || !phi->hasOneUse())
2529 // Replace the increment with a decrement.
2530 BinaryOperator *decr =
2531 BinaryOperator::Create(Instruction::Sub, incr->getOperand(0),
2532 incr->getOperand(1), "tmp", incr);
2533 incr->replaceAllUsesWith(decr);
2534 incr->eraseFromParent();
2536 // Substitute endval-startval for the original startval, and 0 for the
2537 // original endval. Since we're only testing for equality this is OK even
2538 // if the computation wraps around.
2539 BasicBlock *Preheader = L->getLoopPreheader();
2540 Instruction *PreInsertPt = Preheader->getTerminator();
2541 int inBlock = L->contains(phi->getIncomingBlock(0)) ? 1 : 0;
2542 Value *startVal = phi->getIncomingValue(inBlock);
2543 Value *endVal = Cond->getOperand(1);
2544 // FIXME check for case where both are constant
2545 ConstantInt* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2546 BinaryOperator *NewStartVal =
2547 BinaryOperator::Create(Instruction::Sub, endVal, startVal,
2548 "tmp", PreInsertPt);
2549 phi->setIncomingValue(inBlock, NewStartVal);
2550 Cond->setOperand(1, Zero);
2555 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2557 LI = &getAnalysis<LoopInfo>();
2558 DT = &getAnalysis<DominatorTree>();
2559 SE = &getAnalysis<ScalarEvolution>();
2562 // Find all uses of induction variables in this loop, and categorize
2563 // them by stride. Start by finding all of the PHI nodes in the header for
2564 // this loop. If they are induction variables, inspect their uses.
2565 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2566 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2567 AddUsersIfInteresting(I, L, Processed);
2569 if (!IVUsesByStride.empty()) {
2571 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2576 // Sort the StrideOrder so we process larger strides first.
2577 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare(SE));
2579 // Optimize induction variables. Some indvar uses can be transformed to use
2580 // strides that will be needed for other purposes. A common example of this
2581 // is the exit test for the loop, which can often be rewritten to use the
2582 // computation of some other indvar to decide when to terminate the loop.
2585 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2586 // computation in i64 values and the target doesn't support i64, demote
2587 // the computation to 32-bit if safe.
2589 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2590 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2591 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2592 // Need to be careful that IV's are all the same type. Only works for
2593 // intptr_t indvars.
2595 // IVsByStride keeps IVs for one particular loop.
2596 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2598 // Note: this processes each stride/type pair individually. All users
2599 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2600 // Also, note that we iterate over IVUsesByStride indirectly by using
2601 // StrideOrder. This extra layer of indirection makes the ordering of
2602 // strides deterministic - not dependent on map order.
2603 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2604 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2605 IVUsesByStride.find(StrideOrder[Stride]);
2606 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2607 StrengthReduceStridedIVUsers(SI->first, SI->second, L);
2611 // After all sharing is done, see if we can adjust the loop to test against
2612 // zero instead of counting up to a maximum. This is usually faster.
2613 OptimizeLoopCountIV(L);
2615 // We're done analyzing this loop; release all the state we built up for it.
2616 IVUsesByStride.clear();
2617 IVsByStride.clear();
2618 StrideOrder.clear();
2620 // Clean up after ourselves
2621 if (!DeadInsts.empty())
2622 DeleteTriviallyDeadInstructions();
2624 // At this point, it is worth checking to see if any recurrence PHIs are also
2625 // dead, so that we can remove them as well.
2626 DeleteDeadPHIs(L->getHeader());